Wafer Fabrication: Automatic Material Handling System 9783110487473, 9783110486902

This book systematically introduces modeling, performance evaluation and applications of Automatic Materiel Handling Sys

176 83 9MB

English Pages 284 Year 2018

Report DMCA / Copyright

DOWNLOAD PDF FILE

Table of contents :
Preface
Contents
1 Semiconductor wafer fabrication system
2 Automated material handling systems in SWFSs
3 Modeling methods of automated material handling systems in SWFSs
4 Analysis of automated material handling systems in SWFSs
5 Scheduling methods of automated material handling systems in SWFSs
6 Scheduling in Interbay automated material handling systems
7 Scheduling in Intrabay automatic material handling systems
8 Integrated scheduling in AMHSs
9 Scheduling Performance Evaluation of Automated Material Handling Systems in SWFSs
Index
Recommend Papers

Wafer Fabrication: Automatic Material Handling System
 9783110487473, 9783110486902

  • 0 0 0
  • Like this paper and download? You can publish your own PDF file online for free in a few minutes! Sign Up
File loading please wait...
Citation preview

Jie Zhang, Wei Qin, Lihui Wu, Junliang Wang, Youlong Lv, Xiaoxi Wang Wafer Fabrication

Also of interest Productivity and Organizational Management C. Machado, J. P. Davim (Eds.), 2017 ISBN 978-3-11-035545-1, e-ISBN (PDF) 978-3-11-035579-6, e-ISBN (EPUB) 978-3-11-038661-5

Solar Cells and Energy Materials T. Oku, 2016 ISBN 978-3-11-029848-2, e-ISBN (PDF) 978-3-11-029850-5, e-ISBN (EPUB) 978-3-11-038106-1

Innovation Management C. Machado, J. P. Davim (Eds.), 2015 ISBN 978-3-11-035872-8, e-ISBN (PDF) 978-3-11-035875-9, e-ISBN (EPUB) 978-3-11-038667-7

Industrial Software Applications R. Geisler, 2015 ISBN 978-3-11-037098-0, e-ISBN (PDF) 978-3-11-037099-7, e-ISBN (EPUB) 978-3-11-039678-2

Jie Zhang, Wei Qin, Lihui Wu, Junliang Wang, Youlong Lv, Xiaoxi Wang

Wafer Fabrication

Automated Material Handling Systems

Authors Prof. Jie Zhang College of Mechanical Engineering Donghua University Shanghai China [email protected]

Junliang Wang Institute of Intelligent Manufacturing and Information Engineering Shanghai Jiao Tong University Shanghai China [email protected]

Wei Qin Institute of Intelligent Manufacturing and Information Engineering Shanghai Jiao Tong University Shanghai China [email protected]

Youlong Lv College of Mechanical Engineering Donghua University Shanghai China [email protected]

Lihui Wu School of Mechanical engineering Henan University of Technology Zhengzhou China [email protected]

Xiaoxi Wang [email protected]

ISBN 978-3-11-048690-2 e-ISBN (PDF) 978-3-11-048747-3 e-ISBN (EPUB) 978-3-11-048723-7 Library of Congress Cataloging-in-Publication Data Names: Zhang, Jie, 1963 September 21- author. Title: Wafer fabrication : automatic material handling system / Jie Zhang, Wei Qin, Lihui Wu, Junliang Wang, Youlong Lv, Xiaoxi Wang. Description: Berlin ; Boston : Dr Gruyter, [2018] | Includes bibliographical references. Identifiers: LCCN 2018007853| ISBN 9783110486902 (hardcover) | ISBN 9783110487473 (pdf) | ISBN 9783110487237 (epub) Subjects: LCSH: Semiconductor wafers–Automatic control. | Manufacturing processes–Automatic control. Classification: LCC TK7871.85 .Z43 2018 | DDC 621.39/5–dc23 LC record available at https://lccn.loc.gov/2018007853 Bibliographic information published by the Deutsche Nationalbibliothek The Deutsche Nationalbibliothek lists this publication in the Deutsche Nationalbibliografie; detailed bibliographic data are available on the Internet at http://dnb.dnb.de. © 2018 Walter de Gruyter GmbH, Berlin/Boston Typesetting: Integra Software Services Pvt. Ltd. Printing and binding: CPI books GmbH, Leck Cover image: Patrick landmann/science photo library www.degruyter.com

Preface At present, the semiconductor manufacturing industry has become the high-tech key industry that has been promoted and developed in China. With the rapid development of the semiconductor manufacturing technology, wafer sizes have developed from 6 inch, 8 inch, 12 inch, to 18 inch. For a typical 300 mm wafer fabrication system, there are usually 200–600 processing operations for each wafer. As many as hundreds of vehicles are used to fulfil the material handling work. The total delivery distance of each wafer reaches 8–10 miles. Therefore, the automated material handling system with high operating efficiency has been widely adopted in order to improve the wafer processing machine utilization and shorten the chip delivery due date to ensure that wafer manufacturing enterprises have high competitiveness in the market. In the 1990s, well-known experts in the field of wafer fabrication, such as Kumar and Leachman in the United States, had begun to study scheduling problems in wafer fabrication systems. Their results have led to the recent interest in the scheduling problems in wafer fabrication systems. Although there are a lot of methods reported in the literature, the scheduling optimization problems in automated material handling systems have been rarely studied. These scheduling problems are NP-hard problems due to their large-scale, dynamic and real-time features. We have worked on investigating theories and techniques of modeling and scheduling in automated material handling systems in semiconductor wafer fabrication systems. In particular, with the support of National Natural Science Foundation Programs and National High Technology Research and Development Programs, we have published a large number of papers in the field of scheduling in automated material handling systems. This book is a systematic summary of these research results. This book presents methods and technologies of scheduling problems in automated material handling systems comprehensively and systematically. We are grateful to Yinbin Sun and Cong Pan for their assistances in the preparation of this book. Meanwhile, Jungang Yang, Qiong Zhu, Peng Zhang, Xiaolong Yang, Tengda Li and Yaping Zhou completed many auxiliary works, and we thank all of them. We wish to acknowledge a large number of references in the completion of the manuscript. We are sorry for any errors if there is negligence. In addition, we very much appreciate the comments by Prof. Zhengcheng Duan at Huazhong University of Science and Technology, Prof. Yiming Rong at Tsinghua University, and Prof. Guoquan Huang at the University of Hong Kong. The writing of this book has been supported by the National Nature Science Foundation of China (Grant Nos 51435009, U1537110, and 51275307), and by the National Science and Technology Academic Works Publishing Fund. The theories, methods and applications of scheduling for automated material handling systems in semiconductor wafer fabrication systems are rapidly developing and have become a hot topic in the field of scheduling. If you have any questions about shortcomings and mistakes of this book, please do not hesitate to contact us. https://doi.org/10.1515/9783110487473-201

Contents 1 1.1 1.1.1 1.1.2 1.2 1.2.1 1.2.2 1.2.3 1.2.4 1.2.5 1.3 1.3.1 1.3.2 1.4

2 2.1 2.1.1 2.1.2 2.1.3 2.2 2.2.1 2.2.2 2.2.3 2.2.4 2.2.5 2.3 2.3.1 2.3.2 2.3.3 2.3.4 2.3.5 2.4

Semiconductor wafer fabrication system 1 Semiconductor Manufacturing Industry 1 Development and Current Status of the Semiconductor Manufacturing Industry 1 Future Challenges in the Semiconductor Manufacturing Industry 4 Processing of Semiconductor Chips 5 Wafer Preparation 5 Wafer Fabrication 7 Wafer Sort 8 Chip Packaging 9 Chip Final Measurement 10 Constitution of the Semiconductor Wafer Fabrication System Wafer Processing System 11 Material Handling System 14 Production Scheduling in the Semiconductor Wafer Fabrication System 16 References 17

Automated material handling systems in SWFSs 19 Development of Material Handling Systems in SWFSs Semi-automated Material Handling Systems 19 Automated Material Handling Systems 19 Intelligent Material Handling Systems 21 Components of the Automated Material Handling System 23 Carriers 23 Transport System 24 Storage System 26 Tracking System 27 Control System 28 Layout of the Automated Material Handling System Single-spine Configuration 29 Double-spine Configuration 30 Integrated Configuration 30 Perimeter Configuration 30 Mixed Configuration 32 Features of the Automated Material Handling System References 34

19

29

33

10

VIII

3 3.1 3.1.1 3.1.2 3.2 3.2.1 3.2.2 3.3 3.3.1 3.3.2 3.4 3.4.1 3.4.2 3.5 3.5.1 3.5.2 3.6 3.6.1 3.6.2 3.7

4 4.1 4.2 4.3 4.3.1 4.3.2 4.3.3 4.3.4 4.4 4.4.1 4.4.2 4.4.3 4.4.4 4.4.5 4.4.6

Contents

Modeling methods of automated material handling systems in SWFSs 36 Modeling Methods Based on the Network Flow 36 Basic Theory of the Network Flow Model 36 Network Flow Modeling Process for AMHSs 38 Modeling Methods Based on the Queuing Theory 42 Basic Theory of the Queuing Theory Model 42 Queuing Theory Modeling Process for AMHSs 45 Modeling Methods Based on Mathematical Programming The Mathematical Programming Model 50 Mathematical Programming Modeling Process for AMHSs Modeling Methods Based on the Markov Model 54 Basic Theory of the Markov Model 54 Markov Modeling Process for AMHSs 56 Modeling Methods Based on Simulation 58 Basic Theory of the Simulation Model 58 Simulation Modeling Process for AMHSs 61 Modeling Methods Based on the Petri Net 67 Basic Theory of the Petri Net Model 67 Petri Net Modeling Process for AMHSs 68 Conclusion 76 References 77

50 52

Analysis of automated material handling systems in SWFSs 78 Running Process Description 78 Methods for Analyzing Running Characteristics 80 Modified Markov Chain Model 81 Notation and System Assumption 82 MMCM’s State Definition 83 Modeling Process of the MMCM 84 Model Validation 93 Analysis of AMHS Based on the MMCM 99 The Overall Utilization Ratio and the Mean Utilization Ratio of a Vehicle 100 The Mean Arrival Time Interval of an Empty Vehicle 101 Expected Throughput Capability and Real Throughput Capability of an AMHS 101 The Waiting Time of a Wafer 101 Vehicle Blockage-Related Indicators 103 Case Study 104

IX

Contents

4.5

5 5.1 5.1.1 5.1.2 5.2 5.2.1 5.3 5.3.1 5.3.2 5.4

6 6.1 6.2 6.2.1 6.2.2 6.2.3 6.2.4 6.2.5 6.3 6.3.1 6.3.2 6.3.3 6.3.4 6.4

7 7.1 7.2

Conclusion References

105 106

Scheduling methods of automated material handling systems in SWFSs 107 Heuristic Rules-Based Scheduling Methods 107 Heuristic Rules 107 Heuristic Rules Applied in AMHS Scheduling 108 Operation Research Theory-Based Scheduling Methods 111 Operation Research Theory 111 Artificial Intelligence-Based Scheduling Methods 119 Artificial Intelligence Algorithms 119 Intelligent Algorithms Applied in AMHS Scheduling 127 Conclusion 133 References 134 Scheduling in Interbay automated material handling systems 135 Interbay Automated Material Handling Scheduling Problems 135 AMPHI-Based Interbay Automated Material Handling Scheduling Methods 137 AMPHI-Based Interbay Automated Material Handling Scheduling Model 137 Architecture of AMPHI Interbay 139 Vehicle Dispatching in the Interbay Material Handling System 140 Fuzzy-Logic-Based Weight Adjustment 145 Simulation Experiments 153 Composite Rules-Based Interbay System Scheduling Method 162 Global Optimization Model of the Interbay System Scheduling Problem 163 Architecture of Composite Rules-Based Interbay System Scheduling Method 166 Genetic Programming-Based Composite Dispatching Rule Algorithm 167 Simulation Experiments 175 Conclusion 180 References 180 Scheduling in Intrabay automatic material handling systems 182 Intrabay Automatic Material Handling Scheduling Problems GDP-Based Intrabay Material Handling Scheduling Method

182 184

X

7.2.1 7.2.2 7.2.3 7.2.4 7.2.5 7.2.6 7.3 7.3.1 7.3.2 7.3.3 7.4

8 8.1 8.2 8.2.1 8.2.2 8.2.3 8.2.4 8.3 8.3.1 8.3.2 8.3.3 8.3.4 8.4

Contents

Formulation of the Intrabay Material Handling Scheduling Problem 184 Architecture of the GDP-Based Intrabay Material Handling Scheduling Method 186 Fuzzy Logic-Based Dynamic Priority Decision-Making Model 187 Hungarian Algorithm-Based Vehicle Dispatching Approach 190 The Greedy Optimization-Based Vehicle Scheduling Strategy 192 Simulation Experiments 193 Pull and Push Strategy-Based Intrabay Material Handling Scheduling Method 199 Architecture of a Pull and Push Strategy-Based Intrabay Material Handling Scheduling Method 199 VSL and LSV Dispatching Rules Based on Pull and Push Strategy 201 Simulation Experiments 202 Conclusion 203 References 207 Integrated scheduling in AMHSs 208 Description of Integrated Scheduling Problems in AMHSs 208 PMOGA-Based Integrated Scheduling Method in AMHSs 210 Formulation of PMOGA-Based Integrated Scheduling Model in AMHSs 210 Architecture of PMOGA-Based Integrated Scheduling Method in AMHSs 213 Description of PMOGA 213 Simulation Experiments 222 GARL-Based Complex Heuristic Integrated Scheduling Method in AMHSs 226 Basic Rules for Automated Material Handling System Scheduling 227 Construction of the Vehicle’s Path Library 228 Genetic Algorithm-Based Intelligent Routing Selection Approach 229 Simulation Experiments 229 Conclusion 233 References 234

Contents

9 9.1 9.2 9.2.1 9.2.2 9.2.3 9.3 9.3.1 9.3.2 9.3.3 9.4 9.4.1 9.4.2 9.5

Index

Scheduling Performance Evaluation of Automated Material Handling Systems in SWFSs 235 Modeling Requirements of Scheduling Performance Evaluation in the AMHS 235 A Modeling Approach Based on an Agent-Oriented Knowledgeable Colored and Timed Petri Net (AOKCTPN) 236 Definition of an AOKCTPN Model 237 Modeling Process of an AOKCTPN Model 240 Feasibility Analysis of an AOKCTPN 252 Scheduling Performance Evaluation of an AMHS Based on the AOKCTPN 257 Transition Time 257 Scheduling Performance Evaluation Indicators 258 Scheduling Performance Evaluation Method 259 Case Study of Scheduling Performance Analysis of an AMHS 260 Verify the Feasibility of the AOKCTPN Model 260 Implementation of an AOKCTPN 262 Conclusion 268 References 269 271

XI

1 Semiconductor wafer fabrication system 1.1 Semiconductor Manufacturing Industry In the twenty-first century, modern technology experienced rapid development, especially the extensive application of electronic technology, which marked the human society entering the information age. In this “new economy” era, the information industry has become one of the largest and fastest growing industries in the world’s economy. Information and Internet technologies have a revolutionary impact on human economic and social life, while the semiconductor industry is its foundation. Integrated circuits (IC), microwave devices and optoelectronic devices made by semiconductor materials are the core products of the electronics industry and the information industry. They have been widely applied in people’s daily lives and exist all over the world. They have infiltrated into digital entertainment, mobile communications, e-commerce, automotive industry, medical equipment, aerospace and so on. In fact, every step of human society’s development nowadays is dependent on semiconductor chip technology. It can be said that the country that owns the semiconductor manufacturing industry will embrace the future of the world [1].

1.1.1 Development and Current Status of the Semiconductor Manufacturing Industry The semiconductor manufacturing industry is developed on the basis of vacuum tube electronics, radio communications and solid physical technology that emerged in the first half of the twentieth century. After World War II, scientists at Bell Telephone Laboratory began to study the solid-state silicon and germanium semiconductor devices. In 1947, they invented the first solid-state transistor, which led to the solid materials and technology-based modern semiconductor industry. Robert Nois in Fairchild Semiconductor and Jack Kilby in Texas Instruments independently invented the IC in 1959. The concept of an IC is intended to interconnect different components on planar silicon. This concept motivates engineers to design more sophisticated electronic circuits to meet customers’ new demands [2]. With the development of science and integrated technology, the carrier of ICs moved from the era of ordinary silicon material into the wafer era. Semiconductor chips are manufactured through a number of processes in the wafer to create a large number of complementary metal oxide semiconductor components that are connected with a metal wire to become a chip with logic function or storage function. https://doi.org/10.1515/9783110487473-001

2

1 Semiconductor wafer fabrication system

Manufacturing processes are becoming increasingly complex; meanwhile, the semiconductor manufacturing machine costs and semiconductor manufacturing plant construction costs are increasing. More European and American companies have moved out of the semiconductor manufacturing field and focused on design. Semiconductor manufacturing industry enterprises can be classified into three major classes: semiconductor equipment suppliers, semiconductor design company (fabless mode) and semiconductor manufacturing company (foundry mode). Of course, there are some companies, such as Intel Corporation, which set up their own manufacturing plants to produce advanced chips in order to protect the key technology of semiconductor chips. For this kind of manufacturers, its manufacturing process is at least one generation ahead of the manufacturing process of a professional pure semiconductor factory. Wafer manufacturing is the core and foundation of the semiconductor IC industry, which has shown a positive growth trend for a very long time. The Global Semiconductor Consortium (GSA) and the market research firm IC Insights jointly conducted a survey to study the growth trend in the industry. The survey report indicated that the percentage of ICs manufactured by wafer foundries in the entire chip market increased from 21% in 2004 to 24% in 2009 and jumped to 37% in 2014. This suggests that the semiconductor industry is transitioning from Integrated Device Manufacturers (IDM) to fab-site or fabless plant-based models, which is currently in the steep part of the industrial life cycle S-curve. GSA and IC Insights expect sales of ICs manufactured by foundries to

IDM

Foundry $72.1

42%

$70

35% $53.7 25%

$47.9

$50 $42.5 18%

$40

13%

10% $21.6

13%

15%

16% 12%

4%

4%

$20

10% 5% –5%

$10

–15%

$0 2009

2010

2011

2012

2013

2014

2015

2016

2017

2018

Data sources: 2015 Foundry Almanac (IC Insights) *IDM (Integrated Device Manufacturer) Figure 1.1: Forecast of sales and growth rates for global foundries (2015–2018).

Growth

Global sales ($B)

$60

$30

45%

1.1 Semiconductor Manufacturing Industry

3

approach 46% of total industry sales in 2018. Figure 1.1 shows the forecast of sales and growth rates for the global foundries in 2015–2018. Encouraged by demand, foundry companies are also gradually expanding their capacities, and actively developing to become super-factory (MegaFab) models. According to industry consensus, the so-called super factory refers to a 12-inch wafer factory with the monthly production capacity of 100,000–15 million and the amount of $70–$80 billion investment. Semiconductor ICs are applied in almost all electronic equipment. Because of its strong promoting role and large multiplier effect, it is of great significance to computers, household appliances, digital electronics, automation, communications, aerospace and other industries. According to the China Semiconductor Industry Association statistics in 2010–2014, semiconductor IC industry sales in China maintained a double-digit, high growth rate. The proportion of GDP is increasing, as shown in Figure 1.2. Sales revenue data of 2015 must be available now. At present, the output value of the information industry in developed countries occupies 40–60% of the total output value of their national economy, and 65% of the GDP growth is related to the IC industry. Therefore, we can seize the initiative of the national economic development by further developing the IC industry. The semiconductor IC industry in China is also rapidly developing. China has developed a number of highly competitive enterprises including SMIC and Huahong Grace, which have entered the global top ten foundry business (shown in Table 1.1).

Growth

Sales ($B) 3,500 3,000

50.00% 43.30%

3015.4 29.80%

34.30%

2608.51

2,500

30.00% 2158.45

24.30% 2,000

20.00% 20.00%

1933.7 16.20%

1440.15

1,500 1,000

40.00%

1251.3 1246.8 1006.3

10.00%

11.60%

1109.13 0.00%

–0.40%

500

–10.00%

–11.00%

0

–20.00% 2006

2007

2008

2009

2010

Figure 1.2: IC industry sales in China (2006–2014).

2011

2012

2013

2014

4

1 Semiconductor wafer fabrication system

Table 1.1: The global foundry business ranking (2013). Ranking

Company

Country/region

1 2 3 4 5 6 7 8 9 10

TSMC GlobalFoundries UMC Samsung SMIC Powerchip Vanguard Huahong Grace Dongbu TowerJazz

Taiwan, China Milpitas, CA, USA Taiwan, China Korea Mainland, China Taiwan, China Taiwan, China Shangai, China Korea Israel

Sales (100,000,000$)

Year-on-year growth (%)

198.5 42.61 39.59 39.50 19.73 11.75 7.13 7.10 5.70 5.09

17 6 6 15 28 88 23 5 6 –20

1.1.2 Future Challenges in the Semiconductor Manufacturing Industry In the past half century, the semiconductor industry has undergone tremendous changes. In the coming two decades, the semiconductor industry will continue to face dramatic changes and challenges. Three aspects are presented as follows [3]: 1. Technological change is accelerating The semiconductor chip industry began to develop rapidly in 1960, and the smallest feature size of the IC has been drastically reduced in less than 60 years. The size of the chip reduced from 50 μm in 1960 to 0.18 μm in the early twenty-first century. Up to now, the 40 nm IC has been very stable in the market. The 28 nm IC has also officially entered mass production and now it seems to have evolved to 22 nm, 16 nm and even more fine scale. At present, the 22 nm technology in Intel (the world’s most advanced technology company) has entered mass production (leading industry one–two years); 14 nm process is also mature. Rapid technological change leads to the emergence of new products and shortens products’ life cycles. Therefore, the primary challenge of semiconductor manufacturers is to achieve the maximum profit in the increasingly shortened product life cycle by shortening the manufacturing cycle and reducing production costs at a molar speed. 2.

The size of the semiconductor chip manufacturing industry needs to be expanded As the cost of developing and implementing next-generation processes is increasing rapidly, leading companies in the industry must be the forefront of semiconductor process development, and only large companies can afford these costs. Based on this, the acquisition of wafer industry has become the focus of the industry since 2008. Tower Semiconductor Ltd acquired Jazz Semiconductor Inc.

1.2 Processing of Semiconductor Chips

5

in 2008. Global Foundries acquired Chartered Semiconductor in 2009, and became the second largest wafer foundry. NEC and Manulife Semiconductor (leading chipmakers in Japan) have also merged in 2010. This wave of mergers and acquisitions has also affected China. Therefore, the important development trend of the future semiconductor manufacturing industry is to enhance its scale. 3.

Diversification of market demand The dependence on and demand for digital electronic products in modern society will continue to rise rapidly in terms of quantity, quality and diversity. Technological innovation shortens the life cycles of semiconductor products, and fierce market competition reduces the profit in the semiconductor chip unit. In the shortest possible time, enterprises have to meet the dynamic demands of market diversification to produce a large number of new high-quality products to meet customer requirements and to quickly seize the market in order to earn the maximum profits. Therefore, one of the major challenges of semiconductor manufacturing enterprises is to effectively integrate internal and external resources in order to use the cost of less varictics and mass production to meet the demand of multi-species and mass production. The development of the semiconductor industry is of great strategic significance in China; however, it is still in its early stages and faces challenges in technical expertise and has deficiencies in management. The most effective way to grow in the field would be to explore the development path by combining technology and management, which would actively promote the development of the semiconductor industry in China.

1.2 Processing of Semiconductor Chips The entire semiconductor chip manufacturing process consists of five phases: wafer preparation, wafer fabrication, wafer sort, chip packaging and chip final measurement. Among them, wafer fabrication and wafer sort are incorporated as “front-end operations”; chip packaging and chip final measurement are referred to as “back-end operations” (shown in Figure 1.3) [4]. In a semiconductor wafer fabrication system (SWFS), dozens or hundreds of specific IC chips are fabricated on a wafer with silicon as a substrate. The operational risks involved include a large number of workloads, large quantities of expensive processing machines and production steps, and a high degree of re-entrant flows.

1.2.1 Wafer Preparation Silicon is the main semiconductor material used to make IC chips and the most important material in the semiconductor industry. High-purity silicon used to make

6

1 Semiconductor wafer fabrication system

Wafer fabrication

Wafer preparation

Sand

Circuit design

CAD model

Mask

Crystal column Deposition

Coating

Metallization

Cutting

Ion implantation

Wafer sort

Lithography

Etching

Front-end operation

Polishing

Tester

Chip

Chip final measurement

Chip packaging

Back-end operation

Wafer

Figure 1.3: The main steps in semiconductor chip manufacturing process.

the chip is called semiconductor-grade silicon; sometimes, it is referred to as electronic-grade silicon [4]. The basis of wafer preparation is to convert a polycrystalline semiconductor-grade silicon block to a large single-crystal silicon, called a silicon ingot. The head and tail of the monocrystalline silicon rod are cut off and then the rod is mechanically trimmed to the appropriate diameter in order to obtain a “silicon rod” with a suitable diameter and a certain length. Since the silicon is very hard, a diamond saw is used to cut the silicon rod into wafer thin slices of equal thickness. The

1.2 Processing of Semiconductor Chips

7

wafer is then ground in order to reduce the saw marks and damage to the front and back surfaces of the wafer and to thin the wafer. This process is called milling, after which the wafer is cleaned using chemicals, (i.e. sodium hydroxide, acetic acid and nitric acid to clean the surface of the wafer) in order to reduce damage and cracks generated during grinding. Thereafter, in the chamfering process the edges of the wafer are rounded to reduce the possibility of damage during future circuit fabrication. Finally, the wafer surfaces are polished to reduce surface roughness, thereby further reducing the possibility of breakage. The wafer is formed after these steps.

1.2.2 Wafer Fabrication Wafer fabrication is the core of semiconductor chip manufacturing. Wafer manufacturing process is similar to building a house. The wafer is built up layer by layer, and these layers are distinguished by the lithography process. Each layer includes deposition, lithography, etching, ion implantation, electroplating and other processes. The number of process steps varies with the product. In general, the production process of a product consists of at least hundreds of processes [5]. In the wafer fabrication process, the most important feature is its re-entrant characteristic [6]. In 1993, Kumar proposed Re-entrant Lines (unlike Job shop and Flow shop) according to the typical characteristics of the regular reentries of the machining paths in the semiconductor manufacturing system. A product in a reentrant manufacturing system repeatedly revisits the same machine or the same machine group at different processing stages of a process route. 1. Deposition The first step in fabricating a wafer is to deposit a layer of non-conductive silicon dioxide film on the wafer. In the subsequent manufacturing process, the silicon dioxide layer is grown; the deposition process will be carried out many times. The silicon dioxide film forming technique includes physical vapor deposition (PVD) and chemical vapor deposition (CVD). The PVD technique is to apply thermal or kinetic energy to the source of the material to be deposited to decompose it into an aggregate of atoms or atoms, to impinge on the wafer surface, to bond or agglomerate to form a film. The PVD technique consists of three methods: resistance heating deposition method, electron gun evaporation method and sputtering method. The CVD technique is to introduce the reaction gas into the high temperature furnace to produce a chemical effect on the wafer surface with the gaseous chemical raw material, and to deposit a thin film on the wafer surface. 2.

Lithography Lithography is the basis to build up a semiconductor Metal-Oxide-Semiconductor Field-Effect Transistor (MOS) tube and a circuit on a flat wafer, which contains many steps and processes. First, the silicon is coated with a layer of corrosion-resistant

8

1 Semiconductor wafer fabrication system

photoresist, and then light is let through a hollow mask engraved with a circuit pattern to irradiate the wafer. The parts to be irradiated (such as the source and drain regions) are deteriorated, and the gate area is not irradiated and its photoresist remains sticky. The next step is to clean the wafer with a corrosive liquid, and the deteriorated photoresist will be removed to expose the underlying wafers. The gate area is not affected due to protection by the photoresist. 3.

Etching Etching is a technology to remove the material by using chemical reactions or physical impact. Etching techniques contain wet etching and dry etching. Wet etching is to use chemical solutions to etch by chemical reactions, while dry etching is generally plasma etching. Plasma etching may be the physical effect on the surface of the chip by high-energy particles in the plasma, it may be a complex reaction between the active radicals in the plasma and the atoms on the chip surface, or it may be a combination. After the etching process is completed, the photoresist on the wafer surface is removed with ionized oxygen. Then the wafer is cleaned with a chemical reagent.

4. Ion implantation Ion implantation (ion doping) is a technique by which impurities are injected into a specific region of the semiconductor assembly in an ion form to obtain accurate electronic properties. These ions must first be accelerated to a sufficiently high energy and velocity to penetrate (inject) the film to reach a predetermined implant depth. The ion implantation process can precisely control the impurity concentration in the implantation area. Basically, the impurity concentration (dose) is controlled by the ion beam (the total number of ions in the ion beam) and the scanning rate (the number of times the wafer passes through the ion beam). And the depth of ion implantation is determined by the size of the ion beam energy. 5.

Electroplating The function of the chip is achieved by the connection of the transistors. The conductive metal (usually aluminum) is deposited on the wafer surface by a metallization process. Unwanted metals are removed by using photolithography and etching processes to leave circuit metal between connected transistors. Complex wafers need a lot of insulators. A chip must be connected to millions of conductive lines, which include horizontal connections in the same layer and the vertical connections between layers.

1.2.3 Wafer Sort The wafer test results are determined by wafer production and quality verification, also known as die sort or wafer sort. During testing, the electrical capacity

1.2 Processing of Semiconductor Chips

9

and circuit function of each chip are detected. The wafer test includes the following three objectives: first, before the wafer is sent to the packaging plant, the qualified chips are identified. Second, the electrical parameters of the device/ circuit are assessed, and engineers need to monitor the distribution of parameters to obtain the quality of the process. Third, the chip yield is detected. The yield test provides full performance feedback to wafer producers. Qualified chips and defective products on the wafer are recorded on the computer in the form of a wafer chart. In the test, the wafer is fixed on a chuck with vacuum suction. The hair-like probe made of gold wire mounted on the test head is kept in contact with the pad on the grain to test its electrical characteristics. Unqualified grains are marked. When the wafer is cut in units of grain, the marked unqualified grain will be eliminated and will no longer enter the next process so as not to increase manufacturing costs. As the size of the IC continues to increase and the line width shrinks, the circuit density continues to increase. The cost of wafer testing is high and continues to rise because longer test times, more sophisticated power supplies, mechanical devices and computer systems are required to perform test work and monitor test results. The visual inspection system is also more sophisticated and expensive as the chip size expands. The chip designer is asked to introduce the test pattern into the storage array. Test designers are exploring how to make the testing process more streamlined and effective. For example, simplified test procedures are adopted after the chip parameters are evaluated. It is also possible to interlace the chips on the wafer or to test multiple chips at the same time.

1.2.4 Chip Packaging The purpose of chip packaging is to protect the chip from the external environment and to provide good working conditions, which causes the IC to have a stable and normal function. Specifically, the chip and other elements that are arranged in the frame or substrate are fixed and connected by using membrane technology and micro-processing technology in order to constitute the overall three-dimensional structure of the process. The package provides protection for the chip. Non-encapsulated IC chips are generally not directly available. Chip packaging features include: 1. Transmission power, which mainly refers to the power supply voltage distribution and conduction. 2. The transmission of the circuit signal is mainly to minimize the delay of the electrical signal. When wiring, the path between the signal line and the chip and the path through the encapsulated I/O interface should be as short as possible.

10

1 Semiconductor wafer fabrication system

3.

To provide cooling channels, in a variety of chip packaging, distribution of the heat gathered by long-term works of components should be taken into consideration. 4. Structural protection and support, chip packaging for the chip and other connecting parts can provide a solid and reliable mechanical support that can adapt to a variety of working environments and changing conditions.

1.2.5 Chip Final Measurement After the packaging process of the semiconductor chip is completed, the chip package testing is carried out. That is, the structures and electrical functions of the semiconductor components are confirmed so as to ensure that the semiconductor components meet the requirements of the system, which is also known as the final measurement. Package testing includes both quality and reliability testing. Quality inspection is mainly to detect the availability of the chip after packaging, i.e., the chip performance. Reliability testing is to test encapsulation reliability related parameters. The general reliability test project has six parts: pretreatment test, temperature cycling, thermal shock test, high-temperature storage test, temperature and humidity test, and high-temperature boiling test.

1.3 Constitution of the Semiconductor Wafer Fabrication System In the entire semiconductor chip manufacturing process, wafer fabrication is the core stage. Its role is to print multi-layer circuits on silicon or gallium-arsenic compound wafers. In order to achieve this goal, in accordance with a certain pattern, a variety of interrelates involved in the wafer fabrication process are integrated with the interaction of the relevant elements by the wafer fabrication system. According to the physical composition, a wafer fabrication system consists of machine layer, unit layer and system layer (shown in Figure 1.4). The machine layer includes processing machines, buffers, material handling equipment and so on. In general, semiconductor processing machines can be further classified as SPMs (single lot processing machines) and BPMs (batch processing machines). In order to avoid contamination of the wafers during the production process, the wafers are delivered with a standard container in a manufacturing system. In a semiconductor manufacturing system, quantities of wafers are grouped into a standard container, which is called a lot. A lot usually contains 25 wafers. In a wafer fabrication system, a part (e.g., a wafer) may revisit the same machine pool multiple times throughout the manufacturing process. The unit layer of the wafer fabrication system includes an SPM group, a BPM group and a material handling equipment group. The SPM group consists of several

1.3 Constitution of the Semiconductor Wafer Fabrication System

11

Material handling equipment Single lot processing machine Batch processing machine Buffer

Figure 1.4: The physical layout of a wafer fabrication system.

SPMs with the same processing function. The BPM group consists of several BPMs with the same machining function. The material handling equipment group consists of buffer, material handling equipment and other materials handling equipment. The wafer fabrication system is composed of two parts: a wafer processing system and a material handling system.

1.3.1 Wafer Processing System The wafer processing system comprises several main work areas, such as lithography and etching areas. Each area is responsible for completing a certain number of processing operations. Each work area consists of several workstations, equipment group or bay. Each workstation contains several processing machines, work in process (WIP) buffers and transfer mechanisms. In the non-automated wafer processing system, there will be corresponding machine operators. The wafer processing machines in the workstation have substantially the same function, or they can be combined to complete a particular processing step. The entire wafer processing system can be seen as a four-layer architecture consisting of a wafer fab, a machining area, several workstations and equipment, as shown in Figure 1.5. The fab contains four areas: lithography processing zone, etching processing zone, ion implantation processing zone and deposition processing zone. In general, semiconductor processing machines can be further classified as SPMs and BPMs. In a semiconductor manufacturing system, quantities of wafers are grouped into a standard container, which is called a lot. The wafers in a lot are all the same type of products and travel as a unit between machines. Each SPM is capable of performing manufacturing operations on many wafers at a time and processing wafers one by one. After all the wafers in a lot have been completed, this lot of wafers will leave the SPM together. SPMs can be further classified into three classes: single lot – single

12

1 Semiconductor wafer fabrication system

Etching processing zone

Lithography processing zone

FAB Processing zone

Bay 1

Bay 3

Bay 5

Bay 7

Bay 9

Stocker 1

Stocker 3

Stocker 5

Stocker 7

Stocker 9

Stocker 2

Stocker 4

Stocker 6

Stocker 8

Stocker 10

Bay 2

Bay 4

Bay 6

Bay 8

Bay 10

Equipment section

OHT

Equipment

Doped processing zone

Deposition processing area

Lithography

Coating machine

Etching machine

Glue removing machine

Ion implantation machine

Grinding machine

Oxidation furnace

Vapor deposition machine

Legend

Figure 1.5: The four-layer architecture diagram of the wafer processing system.

piece processing machines (Figure 1.6), single lot–multiple pieces parallel processing machines (Figure 1.7) and MPM (single lot–multiple pieces multi-chamber processing machines). MPM can be subclassified as MPM_SC (multi-chamber processing machine with same chambers) and MPM_DC (multi-chamber processing machine with different chambers). For MPM_SC, wafers can be processed in any available chamber, one by one. For MPM_DC, many wafers are loaded, and should be processed in several chambers successively, according to its processing sequence. A typical SPM includes RTP machine, ion implanter, CVD machine, coating and exposure device in lithography process, ion etcher and so on. Each BPM is capable of simultaneously performing manufacturing operations on multiple lots of wafers with the same processing sequence. All the wafers in multiple lots will be processed together, and they will leave the BPM simultaneously. BPMs can be further classified into two classes. The first one is BPPM (batch parallel processing machines) (Figure 1.8), in which multiple lots of wafers are processed in parallel machines. The other one is WPM (wet-bench processing machine) (Figure 1.9), in which multiple lots of wafers are processed in serial machines. Typical BPM includes horizontal and vertical oxidation furnaces, metal-etching machines, wet etching machine and so on.

1.3 Constitution of the Semiconductor Wafer Fabrication System

13

UPS

Figure 1.6: Single lot – single piece processing machine.

Figure 1.7: Single lot – multiple pieces multi-chamber processing machine.

14

1 Semiconductor wafer fabrication system

Figure 1.8: Batch parallel processing equipment.

N2. air box

WTR

Pos. 8 TANK#2

Pos. 7 TANK#3

Pos. 6 TANK#4

Pos. 5 TANK#5

Pos. 4 TANK#6

Process both

Process both

Process both

Process both

Process both

Process both

Pos. 3

Breaker box

Pos. 9 TANK#1

Figure 1.9: Wet-bench processing machine.

1.3.2 Material Handling System The material handling system mainly includes buffer and material handling equipment. The buffer is a temporary storage system of various types. The most common storage system is the automatic shelf, which can be further subclassified as the Cartesian type and the carousel type. Another storage system is UTS (under track storage), which is able to improve the efficiency of AMHS (automated material handling system). The material handling equipment in semiconductor manufacturing systems consists of five types of equipment: vertical lifters, overhead tracks with transport vehicles, RGV (rail-guided vehicle), AGV (automated guided vehicle) and conveyors. The material handling system to transmit wafers in a production unit is called Intrabay AMHS, while the material handling system to transmit wafers among production units is called Interbay AMHS. AGV is widely applied as flexible and intelligent material handling equipment. The AMHS is adopted in wafer fabs because it can be used not only as an automated warehouse for temporary storage of wafers but also as a means for rapid delivering among different work areas. In a highly automated wafer plant, the AMHS connects the relatively independent production units throughout the plant

1.3 Constitution of the Semiconductor Wafer Fabrication System

15

through the transport tracks so as to complete the automatic material transmission of the entire factory. The benefits include: 1. reduction of personnel With the rapid development of wafer processing technology and rapid increase in production, the traditional way requires a large number of material handling personnel. For example, the output of the wafer production line is 40,000 pieces per month, and 200 people are required to carry the material for each shift. The use of the AMHS can greatly reduce the number of staff who are responsible for material handling and improve the accuracy of material transmission, which has become the driving factor for the development of the AMHS. 2.

reduction of labor intensity of operators The wafer container quality of a 200 mm wafer is about 4.5 kg and the wafer container quality of a 300 mm wafer is about 9 kg. Although a person can easily carry such a wafer container once or twice, it will cause repetitive stress damage. Therefore, when converting to a 300 mm wafer processing system, the use of the AMHS can greatly reduce the labor intensity of operators on the shop floor.

3.

wafer damage risk reduction During wafer processing, the total value of wafers in a wafer container is between $10,000 and $1 million. In the traditional manual handling system there is possibility of wafer containers falling, thus causing damage to the wafers. The use of an AMHS can effectively improve the safety of wafer handling.

4. increase in machine utilization Traditional handling methods in a wafer factory can likely cause a loss of 10–20% because the machines are inactive while they wait for material to be delivered. In comparison, an AMHS can effectively shorten the material handling time, thereby improving machine utilization. According to statistics, the use of an AMHS to improve machine utilization can save several billion dollars during the life of the plant. 5.

shortening the production cycle The AMHS can shorten the wafer handling time between machines and provide a set of automatic predictable work flow planning systems, thereby shortening the processing cycle. The processing cycle is a very important indicator for wafer fabs because faster product launches or shorter lead times can significantly affect the value of a product. The efficient performance of an AMHS can effectively improve the wafer factory’s production capacity and improve the production machine utilization. The AMHS itself has a high handling accuracy; it can greatly reduce the production losses due to operator error. The AMHS has particular advantage for 300 mm wafer and above and is more widely used in large-size wafer fabrication plants.

16

1 Semiconductor wafer fabrication system

1.4 Production Scheduling in the Semiconductor Wafer Fabrication System With rapid changes in market demand and intensified international competition, the complexity of production scheduling in semiconductor manufacturing systems lies in the following aspects [7, 8]: 1. Large-scale production In a typical wafer fab, there are dozens of significant processes. Each process consists of 300–900 processing steps on more than 100 hundred machines. The average production cycle time in the production line is approximately 30–60 days by repeatedly implementing manufacturing processes including oxidation, deposition, metallization, lithography, etching, ion implantation, photo-resist strip, cleaning, inspection and metrology. In addition, various types of products are processed in a semiconductor production line at the same time, which are different in thousands of operation sequences; even if similar products have different versions. This large-scale production increases the complexity of semiconductor manufacturing. Furthermore, orders from different customers containing dozens of product types are processed in the semiconductor production line at the same time. Therefore, production scheduling problems become a great challenge in semiconductor manufacturing systems. 2.

Re-entrant flow The nature of semiconductor manufacturing systems is re-entrant. In a traditional manufacturing system such as job shop and flow shop, the re-entrant flow occurs either in individual processes or in a few rework processes, which belongs to the local phenomenon. However, in a re-entrant semiconductor manufacturing system, a part (e.g., a wafer) may re-visit the same machine pool multiple times throughout the manufacturing process. At different stages in the re-entrant flow, a wafer has to compete with others to be processed in the same machine pool. In particular, two reasons make the production scheduling in re-entrant manufacturing systems different from that in traditional manufacturing systems. First, the hierarchical structure of semiconductor products is so strongly related that each layer is processed in similar operation sequences with different materials or precision. Hence, the re-entrant flow is adapted to process the next layer of the same product. Second, since machines required in wafer fabrication are very expensive, there is need to introduce re-entrant flow to improve machine utilization. Because of the re-entrant flow, wafers visit the same machine pool many times throughout the manufacturing process. This results in significant increase in the number of wafers to be scheduled in the same machine pool, which enlarges the solution space and increases the complexity of the production scheduling process.

References

3.

17

Mixed processing mode In general, in semiconductor manufacturing systems, 25 or 50 wafers are grouped into a standard container, called a lot. Lots of wafers are processed in SPM and BPM. One lot of wafers can be processed in an SPM at a time, while lots of wafers with the same serial number or the same name of the manufacturing process can be processed in a BPM at the same time. The maximum batch size is regarded as an important parameter of BPMs, which indicates the maximum processing capacity. As a major feature in semiconductor manufacturing systems, batch processing is an important factor in meeting customers’ demands for delivery due dates, and it affects the total production cycle of the semiconductor production line. Considering differences between SPM and BPM, a different scheduling method should be proposed to properly arrange operation sequences in order to improve the productive efficiency of the total semiconductor production line. Production scheduling in wafer fabrication systems is one of the most complex production scheduling problems [9]. According to the different scheduling objects, production scheduling problems can be classified into three categories: scheduling problems of processing systems, scheduling problems of AMHSs and integrated scheduling problems of processing systems and material handling systems. Production scheduling problems of AMHSs can be further classified as: Intrabay scheduling problem, Interbay scheduling problem and Intrabay and Interbay integrated scheduling problem according to the scheduling scope. The authors have investigated these scheduling problems for over ten years and published some papers. At present, with rapid changes in market demand for semiconductor ICs and the intensification of global economic competition, wafer manufacturing enterprises face enormous pressure to improve production efficiency and reduce production costs. The scheduling methods and techniques of AMHSs with excellent performance are one of the effective and fundamental ways to improve the overall efficiency of AMHSs and wafer processing systems. This book will introduce solutions and techniques of scheduling problems in AMHSs comprehensively and systematically.

References [1] [2] [3] [4] [5]

Huanzi company. China’s integrated circuit industry analysis report for 2011. Beijing, 2011. Wang, Y.Y. Adapt to the need of the industry of the development of microelectronics technology in the 21st century. Physics, 2004, 33(6): 407–413. Jiang, Z.H. The development of integrated circuit industry in China. Proceedings of the 2005 Asia and South Pacific Design Automation Conference, Shanghai, China, 2005, pp. 7–8. Zant, P. Chip Manufacturing, Semiconductor Technology Process and Practical Tutorial (Fourth Edition). Beijing: Electronic Industry Press, 2004. Quirk, M. Semiconductor Manufacturing Technology. Beijing: Electronic Industry Press, 2004.

18

[6] [7]

[8]

[9]

1 Semiconductor wafer fabrication system

Narahari, Y., Khan, L.M. Performance analysis of scheduling polices in re-entrant manufacturing system. Computer Operations Research, 1996, 23(1): 37–51. Uzsoy, R., Lee, C.Y., Martin-Vega, L.A. A review of production planning and scheduling models in the semiconductor industry. Part I: system characteristics, performance evaluation and production planning. IIE Transactions, 1992, 24(4): 47–60. Uzsoy, R., Lee, C.Y., Martin-Vega, L.A. A review of production planning and scheduling models in the semiconductor industry. Part II: shop-floor control. IIE Transactions, 1994, 26(5): 44–55. Kumar, P.R. Scheduling semiconductor manufacturing plants. IEEE Control Systems Magazine, 1994, 14(6): 33–40.

2 Automated material handling systems in SWFSs 2.1 Development of Material Handling Systems in SWFSs In the initial stage of 150 mm and 200 mm semiconductor wafer fabrication systems (SWFSs), material in the workshop is transported and stored manually. Therefore, its transportation process might affect its proper functioning, and can easily lead to incorrect operations, low material handling efficiency, uneven material distribution, idle equipment and long product cycle. With the development of manufacturing technologies and rapid changes of market demands, the traditional manual material handling mode is gradually disappearing from the historical stage. Wafer manufacturing material handling systems have evolved from the stage of semiautomation to automation and are now rapidly moving toward more intelligent ones.

2.1.1 Semi-automated Material Handling Systems After the 200 mm SWFS factory underwent short-term manual operation to fulfill the transportation constraints, storage and distribution of work-in-process (WIP), it then got transferred quickly to semi-automated material handling phase. The material handling in the processing regions (Bay) is developed on the basis of manual operations in combination with the Interbay automated transportation and storage system. However, WIP transportation between machines is still not automated. In order to improve production efficiency, the automated material handling system (AMHS) (shown in Figure 2.1) is adopted in many 200 mm SWFS factories. It is a semi-automated material handling system, i.e., the first generation of an AMHS [1]. In the first generation of an AMHS, WIP is intuitively controlled and manual transportations without any added value are substituted. The storage capacity and transfer speed of an AMHS must meet production demands and ensure that operators obtain the required material on time. In a 200 mm SWFS factory, the AMHS has played a very important role, which helps implement mass production in semiconductor factories and improve production efficiency and shorten products’ cycle times.

2.1.2 Automated Material Handling Systems With the emergence of 300 mm semiconductor factories, the automated material system (i.e., the second generation of an AMHS) came into being. It has been greatly improved on the basis of the original semi-automated material handling system. https://doi.org/10.1515/9783110487473-002

20

2 Automated material handling systems in SWFSs

Bay 1

Bay 3

Stocker

Stocker

Processing machine

Bay 5

Stocker

OHS Interbay material handling system

OHS Stocker

Stocker

Stocker

Bay 2

Bay 4

Bay 6

Figure 2.1: An AMHS in 200 m semiconductor manufacturing factory.

First, the Intrabay delivery system is added to the existing Interbay system. Accordingly, additional features such as transfer interface between Interbay and Intrabay systems are added to the stockers (shown in Figure 2.2). In addition, the stocker capacity is increased for the Intrabay WIP storage. In the second generation of an AMHS, a new vehicle – overhead hoist transport (OHT) – is utilized to deliver cassettes directly to the processing machines. The cassette is first moved from the Load Port of the processing machine to the OHT and then transported to the destination stocker in another bay. It is then moved to the Load Port of the Interbay by the robot of the destination stocker and then loaded to the overhead shuttle (OHS). Then, the cassette is moved to the stocker which is the nearest machine for the next operation. When this cassette is to be processed, the OHT in Intrabay automatically loads and transports this cassette to the destination processing machine and places it at the Load Port of the processing machine. The entire delivering sequence is shown in Figure 2.2. The second generation of an AMHS is a highly mechanized and segmented material handling system. The entire delivery process consists of two types of transportation to stockers and OHT only works in the specified Intrabay. Each segmented part (Intrabay, stocker, Interbay) is united by the central control system. Transport management is confined to the area related to transport operations and the transfer speed of the entire system is restricted. This AMHS design is established on the basis of the early process flow of the semiconductor factories, in which the system should assign OHTs to each Intrabay according to its production rate. The flexibility of an AMHS is limited by the segmented layout, which is unable to respond quickly to changed process or other abnormal events. In order to achieve the goal of the entire manufacturing system, segmented 300 mm AMHS can respond according to the demands of the manufacturing

2.1 Development of Material Handling Systems in SWFSs

Bay 1 (1) (2)

Bay 3

Bay 5

Processing machine

OHT

OHT OHT

(3)

Stocker

Stocker

Stocker

OHS OHS (5)

(4) Stocker

Interbay material handling system (7)

Stocker

(8)

Stocker

(6)

OHT (9)

OHT

OHT Bay 2

Bay 4

21

OHT

(1) Pick-up cassette (2) Transport in intrabay (3) OHT to stocker (4) Stocker to OHS (5) Transport between bays (6) OHS to stocker (7) Stocker to OHT (8) Transport in intrabay (9) Drop-off cassette

Bay 6

Figure 2.2: The segmented AMHS in a 300 mm semiconductor factory.

execution system (MES). However, the number of OHSs in Interbay and the number of OHTs in Intrabay are fixed. Reconfiguration of the vehicles should be accomplished manually, which restricts the AMHS’s response speed to the changes of the plant. In this design, the stocker not only plays a role in storing wafers, but also plays a more important role of acting as a bridge in delivering wafers. Therefore, the stability of stockers, to a large extent, will affect the production capacity of all machines in this bay. For example, when there is a breakdown of the stocker of a bay (i.e., it does not work properly), the wafers to be processed in this bay cannot be sent to the specified processing machines and the processed wafers cannot be delivered to next machine on time.

2.1.3 Intelligent Material Handling Systems On the basis of the 300 mm automated material handling system, equipment suppliers and solution providers of an AMHS apply various advanced technologies and try to further improve the agility and flexibility of AMHSs toward the intelligent direction. In the embryonic intelligent AMHS, an OHT system is extended to the Interbay area, in which OHTs are able to roam anywhere in the factory. Thus, the wafer cassette will not be temporarily stored in the stocker when it is delivered between different bays. Instead, the wafer cassette is directly delivered to the processing machine, which takes a shorter delivering time. The entire material handling system can achieve dynamic load balancing and “predict” required production rate and allocate proper number of vehicles in each bay. Tasks are dynamically allocated to the nearest vehicle: when a vehicle completes a task, the system will rationally assign

22

2 Automated material handling systems in SWFSs

it another task based on its location. Even if this task is assigned to a farther away vehicle at this time, it is possible to adjust the assignment [2]. In the intelligent AMHS, the storage location of wafer cassettes is shifted from stockers to a place nearest to next processing machine, which is called under-track storage (UTS) or zero-footprint storage (ZFS). UTS uses part of the reserved space in the clean room and thus will not take up the space of processing machines (shown in Figure 2.3). UTS has the following significant advantages: (1) it places wafer lot closer to processing machines in order to shorten the delivering time from processing machines to stockers; (2) it minimizes the footprint of stockers in order to save space for the clean room to place more processing machines; (3) it has cut down the cost to a large extent. UTS not only can reduce the fixed investment such as stockers, but also can improve the reliability of equipment in order to save the cost of regular maintenance. To support the integrated AMHS, dynamic routing selection and UTS storage, the transportation control system and the material control system of an AMHS should have intelligent functions such as correct allocation of tasks, efficient utilization of resources, dynamic balance of fab load, continuous detection of hundreds of parameters associated with WIP and timely responses. Through the unity and integration of Interbay and Intrabay systems, the control system can visually monitor and coordinate each delivery step. By dynamically controlling the position of OHTs, the resources within the material handling system are fully optimized and utilized. The dynamic decision function enables the AMHS to transfer from the design concept which shortens cycle time by rapid mechanical motion in the early stage to the new design concept which controls the vehicle’s location intelligently and allocates tasks properly.

Bay 3

Bay 1

Processing machine

Bay 5

OHT

OHT

OHT OHT OHT

Stocker or carousel

Interbay material handling system Stocker or carousel

OHT OHT OHT Bay 2

OHT Bay 4

OHT UTS

Bay 6

Figure 2.3: An AMHS with UTS in the 300 mm semiconductor factory.

2.2 Components of the Automated Material Handling System

23

In addition, different information systems and automated handling systems can be integrated together by a fab management software package. Additional values and benefits are obtained when these systems are pre-integrated and their advantages are fully realized. By this kind of integration, the demands of material can be expected by an AMHS, and can be dynamically adjusted according to changes in the manufacturing system. The current 300 mm AMHS can achieve the function which cannot be achieved by traditional fab and the early AMHS, which brings more added values to the semiconductor wafer fab.

2.2 Components of the Automated Material Handling System The AMHS in the semiconductor wafer fabrication system consists of five key important components: carriers, transport system, storage system, tracking system and control system.

2.2.1 Carriers The wafer carrier is a wafer cassette in the transport process, which is also known as wafer lot. In the automated material handling system, the wafer carrier as a unit is transported by the vehicle, and usually its capacity is limited to 25 wafers. One vehicle can transport only one to two wafer carriers. In both of the early 200 mm AMHS and the current 300 mm AMHS, the design of wafer carriers is related to whether it can automatically transport wafers, which is also the basis of the hardware of an AMHS and a key step to optimize the AMHS. As there is development in wafer fabrication technologies and as the wafer size becomes larger, the size and structure of wafer carriers are also changing, which mainly has the following types. 1. Standard enclosed wafer carrier In the 200 mm wafer factory, the standard mechanical interface (SMIF)-enclosed wafer carrier is commonly adopted. In each processing machine, there is a standard port to open the wafer carrier. In addition, in the factory adopting SMIF, each machine has a micro environment, which provides a clean environment when a wafer is moved out of the wafer carrier. The SMIF wafer carrier has a universal automation interface which prevents wafers from the potential contamination in the factory; therefore, the SMIF wafer carrier is ideal for the AMHS. 2.

Opening cassette The wafer in the opening cassette is exposed to the surrounding environment. Thus, with respect to the factory adopting SMIF carriers, the factory using opening

24

2 Automated material handling systems in SWFSs

cassettes should be kept cleaner. Opening cassettes could be transported separately and also be stored in a sealed box before transportation. For opening cassette and storage boxes, one major drawback is that the automation is not taken into consideration. Therefore, the interface and the robot must be customized additionally in order to move the opening cassette into or out of the automated material handling system. 3.

Front opening unified pod In the transition to 300 mm wafer, the focus of the worldwide factories is on the standardization of the wafer carrier. Eventually, an integrated 1wafer carrier with a friendly microenvironment and automation ready, which is called the front opening unified pod (FOUP), is selected as the industry standard, as shown in Figure 2.4. Compared with 200 mm SMIF wafer carrier, 300 mm FOUP seals wafers within a controllable environment. Moreover, the robot handling SMIF wafer carriers can also handle FOUPs. There are two differences between FOUP and SMIF: (1) FOUP opens a door at the front, while SMIF wafer carriers open a door at the bottom; (2) FOUP has an embedded wafer holder inside, while SMIF wafer carriers has a removable wafer holder inside.

2.2.2 Transport System Material movement within the AMHS is usually fulfilled by the transport system which consists of various transport vehicles, rails and control system. Transport vehicles generally include: OHT, OHS, AGV, RGV and conveyors. 1. Overhead hoist transport As shown in Figure 2.5, the OHT runs along the track mounted on the ceiling and these tracks are connected to the pick-up/drop-off ports of processing machines. The OHT can lift wafer carriers to these pick-up/drop-off ports. The

Figure 2.4: Front opening unified pod.

2.2 Components of the Automated Material Handling System

25

Figure 2.5: Overhead hoist transport (OHT).

Figure 2.6: Overhead shuttle (OHS).

OHT system is used in both the 300 mm and 200 mm SMIF factories. One major advantage of OHTs is that it does not require any floor space in the factory. Another important advantage is that it can adapt flexibly to changes in the plant layout. 2.

Overhead shuttle Similar to OHT, the OHS also runs along the track mounted on the ceiling, as shown in Figure 2.6. The OHS is responsible for delivering wafer carriers between stockers. Like the transport vehicles that are used between bays, the OHS is also applied in both 200 mm and 300 mm SMIF factories. An OHS can transport one or two wafer carriers and often completes transport tasks together with OHT, RGV and AGV. There are two kinds of OHS: one uses its own robot to move the wafer carrier from the stockers and send to the OHS, and the other uses the robot of the stocker to load wafer carriers.

3.

Rail-guided vehicle The RGV is a transport vehicle used on the ground and can be used to handle material within and between workshops. The RGV is one of the fastest

26

2 Automated material handling systems in SWFSs

material handling vehicles and thus is commonly used in workshops that demand the highest production rate. For security reasons, the RGV cannot coexist with factory personnel, which becomes one of its limitations. Another limitation is that RGV cannot support the flexibility of changes in the factory layout. Although RGV has been used in the 200 mm SMIF factory, it is gradually being replaced by OHT in the 300 mm factory. 4. Automated guided vehicle The AGV is also a transport vehicle that operates on the ground and can reach anywhere in the factory by program control. The AGV can be used for material handling within and between workshops. The AGV is slower than other material handling vehicles but it adapts flexibly to changes in the factory layout. Under certain conditions, the AGV can temporarily coexist with workers in the same workshop. The AGV is mostly used in 200 mm SMIF factories and is rarely used in 300 mm SMIF factories. 5.

Conveyor The conveyor is placed in the ceiling and can send 200 mm and 300 mm wafer carriers. Although the conveyor system is slower than other material handling systems, it provides high material throughput because the movement of wafer carriers will not be influenced by the number of vehicles. The conveyor is usually used in the point-to-point transport, such as transporting wafer carriers between segmented factory buildings. The conveyors in combination with OHTs are taken into consideration to be adapted in high material throughput areas.

2.2.3 Storage System In the process of wafer fabrication, specified storage equipment and its corresponding management system are designed to store wafers. Currently, there are two types of storage system: stocker and under-track storage (UTS) system. The stocker is a storage system that is used to store large number of wafer carriers. A stocker consists of one or more input/output ports, shelves for placing wafer carriers and a robot for handling wafer carriers between shelves. The shape of stockers varies and its capacity ranges from 50 to thousands of wafer carriers. In the factory, there are two kinds of stockers: descartes type and carousel type. There is only one robot in descartes stocker, which can run in three dimensions inside the system and move the wafer carriers from the input/output port to the vertical shelves. Descartes stockers are popular in 200 mm and 300 mm wafer factories. The wafer carriers in carousel stockers are placed on horizontal or vertical shelves and the robot moves in one direction inside the system.

2.2 Components of the Automated Material Handling System

27

Therefore, the carousel stockers are placed more densely and are more reliable than descartes stockers. In the 300 mm factory, carousel stockers and OHTs are combined for automation applications. UTS is also a common method for wafer storage, which can be combined with OHTs. UTS contains a shelf under the track of OHTs, which can store wafer carriers, as shown in Figure 2.7 UTS has three outstanding advantages: (1) it does not occupy floor space; (2) it is highly reliable and (3) it improves the transport efficiency of an AMHS. Because the wafer carrier can be stored beside the next processing machine, the time to store wafer carriers in stockers is reduced. In addition, UTS provides more input/output ports instead of two ports in stockers, which alleviates the congestion of OHTs and improves the efficiency of an AMHS.

2.2.4 Tracking System In modern 300 mm wafer fabs, the number of wafer WIPs is up to 30,000 to 50,000 pieces and each wafer typically has 200–600 machining processes. The transportation distance to complete all the processes is up to 8–10 miles. Faced with such a large-scale material handling, it is necessary to adopt tracking systems to trace the wafer carriers. The tracking system consists of hardware-level identification (such as RFID, two-dimensional code, etc.), reading devices and software-level management systems. In recent years, RFID technology has been developing rapidly. Because the tags can be read without sight, RFID technology has become the mainstream for

Figure 2.7: Under track storage system.

28

2 Automated material handling systems in SWFSs

tracking wafers. In IBM’s semiconductor manufacturing factory located in New York, RFID technology is used to track thousands of orders of chips. The factory covers an area of 140,000 square feet and has equipment worth billions of dollars. It is the first fully automated semiconductor manufacturing factory and operates since mid-2002. In its tracking system, each wafer carrier is customized according to the 300 mm wafer standard, is made of polycarbonate and is implanted a passive RFID tag with unique identification code which can be read by a RFID reader spread over the processing machines. Each processing machine contains a RFID reader, a lithographic exposure tool, metal deposition equipment, chemical mechanical polishing, and an oven, and involves processes such as chemical etching and ion implantation. The reader can also be overall planned and distributed in the device of material handling systems. The reading range of thousands of RFID readers is 3–6 inches. The tracking system in the factory can trace every wafer, including the specific locations where they are being processed. By integrating RFID technology in the production tools, the manufacturing system can determine which product to process next, how to process and when to process.

2.2.5 Control System The control system in an AMHS consists of stocker controller (STC), transport controller (TSC) and auxiliary control software. At different levels, the control system is divided into equipment control system and material control system. The former is used to manage vehicles, stockers and other hardware devices and responsible for their proper operations. The latter is used to guarantee the on-time delivery of wafers and the transport planning of vehicles, including material control software, scheduling and dispatching software, and so on. 1. Material control system (MCS) The MCS is a real-time central control system and it is mainly responsible for the management and control of products, masks, and other transport and storage material. The MCS translates the instructions of operators and other software to the commands that material handling device can understand. In addition, the MCS has a database to identify and track material in the factory. In order to optimize the factory, the MCS must include algorithms dealing with advanced tracking, merging, space splitting, prevention of material shortage and pollution. In order to adapt to continuous changes in the factory, it should have the abilities of being dynamically expanded and repeatedly arranged. 2. Scheduling and dispatching software The scheduling and dispatching system will analyze the status of the wafer fabrication system and trigger material movement in the MCS system, which

2.3 Layout of the Automated Material Handling System

29

can improve on-time delivery rate and reduce WIP. These software packages review the current status of the factory and continuously decide when to start and terminate products and how to balance demands of customers’ shipments to minimize the inventory holding cost. Many rules are applied in scheduling and dispatching software of wafer material handling in order to optimize the efficiency of material handling system and improve on-time delivery rate, and also reduce inventory levels, production cycle times and its fluctuations.

2.3 Layout of the Automated Material Handling System The layout design is an important part in the design phase of the automated material handling system, which aims to allocate all kinds of production processes and related resources to reasonable positions. Under this circumstance, all the WIP can flow with the highest efficiency in the factory, which reduces capital and product costs and shortens the cycle times. In terms of the topological structure, layouts commonly adopted in current AMHSs can be classified as follows [3–8].

2.3.1 Single-spine Configuration The single-spine configuration is the general layout for 300 mm fabs. In this layout, a central aisle is connected with several machines installed in several bays. As shown in Figure 2.8, stockers are located at the intersection of the main aisles of

Bay Stocker

1

10

3

6

4

Turntable Travel direction

Tool–tool direct transport

AMHS 7

2

5

Figure 2.8: Spine configuration for fab layouts.

9

8

30

2 Automated material handling systems in SWFSs

each bay. An AMHS based on a spine configuration consists of one directed flow loop and several crossover turntables to change the travel directions.

2.3.2 Double-spine Configuration In some overhead material handling systems, a double-spine configuration is also adopted, as shown in Figure 2.9.

2.3.3 Integrated Configuration The traditional segmented AMHS is highly automated, but wafer delivery among different production regions must be completed via stockers. This way not only affects the transportation efficiency, but also brings disastrous effects to the critical processing machines when stockers break down. Therefore, an integrated material handling system comes into being, in which inter-regional transport for key processing machines can be completed via OHT, i.e., Tool– Tool direct transport. The feature of this design is that wafer transport between key processing machines does not rely on transit of stockers in order to improve the efficiency of an AMHS. As shown in Figure 2.10, in the factory without Tool–Tool direct transport, the transport route from Tool A to Tool B is A→Stocker1→Stocker4→Tool B. But in the factory with Tool–Tool direct transport (shown in Figure 2.11), the transport route from Tool A to Tool B is Tool A→Tool B, which is achieved through inter-regional transport by OHT.

2.3.4 Perimeter Configuration The perimeter configuration is another widely applied configuration. As shown in Figure 2.12, there are two physically separated loops in this layout. Several crossover turntables are adopted to switch the direction loops. Feindel and Kaempf suggested connecting the Interbay with Intrabay material transportation facilities by using a conveyor-type continuous flow transport (CFT)

X-spine

Y-spine

Figure 2.9: Double-spine layouts.

2.3 Layout of the Automated Material Handling System

Bay 1

Bay 3 Tool A

OHT

Bay 5

OHT

OHT

Stocker 3

Stocker 1

31

Stocker 5

OHT OHT Stocker 2

Stocker 4

OHT

Tool B

OHT

Bay 2

Stocker 6

OHT

Bay 4

Bay 6

Figure 2.10: Segmented AMHS without Tool-Tool handling ability.

Bay 1

Bay 3 Tool A

Bay 5

OHT

OHT OHT

OHT

OHT

OHT

OHT

Stocker 2

Bay 2

Stocker 4

OHT

Tool B

OHT

OHT

Stocker 3

Stocker 1

OHT

OHT

Bay 4

Bay 6

Figure 2.11: Integrated AMHS with Tool-Tool handling ability.

facility in order to reduce the MH time. The requirements of large end-of-bay stockers were reduced by using local buffering facilities. An overhead or floor-mounted conveyor was proposed for high-capacity Interbay with long-distance and point-to-point pod transport. It is suitable for small zero-footprint buffers close to processing tools. The standard deviation caused by irregular MH equipment availability was eliminated, and it was possible to obtain more predictable delivery times. As shown in Figure 2.13, a buffer loop was used to provide space for WIP to wait until a tool became available. A robotic hoist or elevator was proposed to transfer pods between the CFT buffer and the tool load port.

32

2 Automated material handling systems in SWFSs

Bay AMHS

8 1

2

Travel direction

7

9 3 Turntable

10 5

6

4

Stocker

Figure 2.12: Perimeter configuration for fab layouts.

Material flow turntables

Buffer loop

TOOL

Elevator or robotic hoist

Overhead conveyor system Material flow

Figure 2.13: Plan view of zero footprint CFT buffer loop.

2.3.5 Mixed Configuration Some scholars and engineers proposed a configuration that is composed of two Interbay track loops, one is the spine type and the other is the perimeter type, as shown in Figure 2.14.

2.4 Features of the Automated Material Handling System

Bay

Bay

Bay

Bay

33

Stocker

Feasible flow direction Spine (interior) Guide-path loop Shortcut Perimeter (exterior) guide-path loop

Bay

Bay

Bay

Bay

Figure 2.14: An example of a two-row dual-loop bay layout.

2.4 Features of the Automated Material Handling System The characteristics of an automated material handling system are summarized as follows: 1. Large-scale. In a typical 300 mm wafer fabrication system, there are usually hundreds of processing machines. Each cassette has 200–600 processing operations. The total delivery distance of each cassette reaches 8–10 miles. 2. Stochastic and dynamic. In the Interbay material handling system, a number of vehicles are assigned to deliver different cassettes in a same period. Therefore, the moving requests normally arrive at the system in a random way and unexpected blockage of vehicles often arises. The uncertainty of the vehicle location and status significantly increases the complexity of the scheduling problem. 3. Multiple-objective. Normally there are two types of indicators used to measure the performance of an Interbay material handling system: the measurements for transportation and manufacturing performance. In practical industrial manufacturing, the production scheduling process of Interbay material handling is required not only to increase the transport efficiency but also to meet the production strategy. The balance of various objectives needs to be considered in scheduling Interbay material handling. The measurements for transportation performance mainly refer to: (a) mean no-load arrival time of vehicles; (b) mean transport time of vehicles; (c) mean waiting time of cassettes; (d) mean delivery time of cassettes and (e) mean utilization of vehicles. The measurements for manufacturing performance include: (a) throughput of the

34

2 Automated material handling systems in SWFSs

Chapter 1: Semiconductor wafer fabrication systems

Chapter 2: Automated material handling systems in SWFSs

Chapter 3: Modeling methods of automated material handling systems in SWFSs

Running characteristics analysis of AMHSs

Chapter 4: Analysis of automated material handling systems in SWFSs

Scheduling of AMHSs Chapter 5: Scheduling methods of automated material handling systems in SWFSs Chapter 6: Scheduling in Interbay automated material handling systems Chapter 7: Scheduling in Intrabay automated material handling systems Chapter 8: Integrated scheduling in automated material handling systems in SWFSs

Performance evaluation of AMHSs

Chapter 9: Scheduling performance evaluation of automated material handling systems in SWFSs

Figure 2.15: Organizations of this book.

manufacturing system; (b) mean production cycle time; (c) utilization of key equipment and (d) average due date satisfaction rate. Due to these features, it has become a huge challenge to model, analyze and schedule automated material handling systems in semiconductor wafer fabrication systems. In this book, on the basis of Chapter 1, this chapter introduces SWFSs and AMHSs, and the following chapters will address modeling, running characteristics analysis and performance evaluation methods of AMHSs. The organization of each chapter is shown in Figure 2.15.

References [1] [2]

Van Antwerp, K. Automation in a semiconductor fab. Semiconductor International, December, 2004 [Online] Available http://www.semiconductor.net/article/CA483805.html. Kondo, H., Harada, M. Study for realizing effective direct tool-to-tool delivery semiconductor manufacturing. ISSM 2005, IEEE International Symposium. 2005(13): 21–24.

References

[3] [4] [5] [6] [7]

[8]

35

Wiethoff, T., Swearingen C. AMHS software solutions to increase manufacturing system performance. Advanced Semiconductor Manufacturing Conference. 2006(17): 306–311. Zhao, W., Mackulak, G.T. Reducing model creation cycle time by automated conversion of a CAD AHMS layout design. Simulation Conference Proceedings in 1999 Winter. 1999(1): 779–783. Kaempf, U. Automated wafer transport in the wafer fab. Proceedings of the IEEE/SEMI Advanced Semiconductor Manufacturing Conference. 1997: 356–361. Lin, J.T., Wang, F.K., Wu, C.K. Connecting transport AMHS in a wafer fab. International Journal of Production Research, 2003, 41(3): 529–544. Montoya-Torres, J.R. A literature survey on the design approaches and operational issues of automated wafer-transport systems for wafer fabs. Production Planning & Control, 2006, 17(7): 648–663. The International Technology Roadmap for Semiconductors (ITRS), 2004 edition, SIA Semiconductor Industry Association. 2006, http://www.itrs.net.

3 Modeling methods of automated material handling systems in SWFSs The characteristics of an automated material handling system (AMHS) in semiconductor wafer fabrication systems (SWFSs) are large-scale, real-time, stochastic and multi-objective. In order to design an effective scheduling method to optimize the wafer material handling and production process, it is necessary to establish a reasonable system model that is an abstract description of the system. Modeling is an important means and prerequisite to analyze and make decisions for the system [1]. Specifically, we hope to use the established model to reveal the process and mechanism of inherent operational change in an AMHS to determine a variety of factors (e.g., the number of vehicles, the number of shortcuts, rail layout, vehicle speed, vehicle jam probability, wafer card waiting time distribution) on operation performances. At the same time, with the model of the AMHS, the performance of a scheduling method can be evaluated quickly and effectively in order to verify the effectiveness of the scheduling method. At present, modeling methods of wafer manufacturing include network flow model, queuing theory model, mathematical programming model, Markov model, simulation model and Petri net model.

3.1 Modeling Methods Based on the Network Flow The network flow theory is a kind of theory and method in the graph theory. It primarily studies a class of optimization problems on networks. In 1955, T.E. Harris first proposed the problem of seeking maximum traffic volume between two points on a given network when studying the maximum railway throughput. L.R. Ford and D.R. Fulkerson et al. proposed an algorithm to solve these problems, and established the network flow theory [2].

3.1.1 Basic Theory of the Network Flow Model The graph G is an ordered pair (V, E), where V is regarded as the vertice set, E is the edge set, and E and V do not intersect. They can also be written as V (G) and E (G). The directed graph is used to represent the ordered binary, and the direction of each arc in the directed graph is given. In general, an arc can be represented as an ordered pair (i, j), which points to node j from node i. Arrows are used to specify the orientation, as shown in Figure 3.1. One common representation is the node-arc incidence matrix of a graph. The structure is constructed by listing the nodes vertically and arcs horizontally. Then https://doi.org/10.1515/9783110487473-003

3.1 Modeling Methods Based on the Network Flow

37

Figure 3.1: Directed graph.

below the arc in the column, a “1” is at the position of the corresponding node i, and a “–1” is at the position corresponding to node j. The incidence of Figure 3.1 is represented as follows: (1,2) (1,4) (2,3) (2,4) (4,2) (4,3) 0 1 1 2B B −1 3@ 4

1

1 1 −1

−1

1

−1

−1

1

C C 1 A −1

Suppose that V = fjv1 , . . . , vm g and E = fe1 , . . . , en g. Introduce an m × n node-arc incidence matrix A, and its elements aij are represented as follows: 8 < 1, aij = − 1, : 0,

node vi is the starting point of the arc ej ; node vi is the ending point of the arc ej ; otherwise .

There may be a stream along the arc, and a stream in a directed arc (i, j) is denoted as xij ≥ 0. The flow must meet the preservation requirements at each node in the network. That is, the traffic volume of node i is n X

xij −

j=1

n X

xki = bi

k=1

The first term on the left is the sum of the flows flowing out of node i, and the second term is the sum of the flows entering node i; bi is the storage requirement of node i. Of course, if the arc from node i to node j does not exist, xij does not exist either. In particular, the traffic volume can neither be created nor be lost at the node unless the node is a starting point or an ending point, i.e., the total traffic volume entering a node must be equal to the total traffic volume flowing out of the node. Therefore, n X j=1

xij −

n X k=1

xki = 0

38

3 Modeling methods of automated material handling systems in SWFSs

The first term on the left is the sum of the flows flowing out of node i, and the second term is the sum of the flows entering node i. Many practical network flow problems such as transportation problems, distribution problems, shortest path problems and maximum flow problems can be transformed into such network flow planning problem.

3.1.2 Network Flow Modeling Process for AMHSs The method based on the network flow model usually simplifies the material handling problem in AMHSs to a deterministic directed graph model. The material handling time is simplified to a definite time. The long-term behavior of the directed graph model is analyzed statistically, and then the network flow planning problem is solved. In general, an available wafer fabrication workshop is divided into equal-sized unit-length squares (workstations) by using a space filling curve (SFC) method, where each machining area consists of several workstations. In Figure 3.2, the width of each workstation is the height of the rectangle. Each machining area is assigned a unique value as its address. The address numbering sequence is the same as the direction of the material flow in the AMHS, and the box from the upper-left corner serves as the starting point. The position of the loading/unloading point (P/D) of the machining area is indicated by the boundary distance from the starting workstation of the machining area. Therefore, the P/D position of the upper row of the machining area is represented by the distance between the P/D point and the left side of the starting rectangle. And the P/D position of the lower row of the machining area is represented by the distance between the P/D point and the right side of the starting rectangle. Obviously, the P/D point is close to the material handling system. Based on the SFC method, it is first necessary to develop an initial flow sequence to construct a layout, as shown in Figure 3.3. Once the spatial fill curve and the initial flow order are determined, the design of the material handling system becomes a matter of determining the number of shortcuts required in the layout and each shortcut location. The goal of this model is to find the optimal solution, which minimizes the increase in the shortcut costs while minimizing the reduction in handling costs.

H1

1

4

5

6

H2

12 11 10 9

8

7

2

3

Address assignment

AMHS Unit block

Figure 3.2: Address assignments for spine configuration.

39

3.1 Modeling Methods Based on the Network Flow

Upper row

4

2

SFC

AMHS Lower row

3

1

Flow sequence: 2 4

Unit block 1 3

Figure 3.3: SFC example for spine configuration.

Crossed turntable Machining area P/D position

0

2

3

1

1.5

e

e

1

4

5.5 5.0

Figure 3.4: P/D point sorting example.

Machining area P/D point

1

10

3

4

6

Crossed turntable Travel direction

Tool–tool direct transport Flow direction 7

2

5

9

8

Figure 3.5: Spine configuration integrated design.

The position of the crossed turntable is determined. First, from the left border, the P/D point is placed onto the horizontal axis. The two-dimensional coordinates of the P/D point are represented by a single dimension, as shown in Figures 3.4 and 3.5. In this example, there are three optional positions available for setting the crossed turntable: one between machining area 2 and machining area 3, one between machining area 3 and machining area 1 and one between the machining area 1 and machining area 4.

40

3 Modeling methods of automated material handling systems in SWFSs

If the vertex set (V) represents the P/D point of the machining area and the crossed turntable, and the edge combination (E) represents the directional flow curve between the vertices, then the standard network representation G = (V, E) indicates the problem. Figure 3.5 shows the spine configuration integrated design. A feasible layout of the AMHS and a candidate shortcut scheme are considered. Each arc has the same length. δij is denoted as the directional flow distance of arc(i, j). The directional flow density between the P/D points of the machining area is k; then the set of directional flow density is defined as K, k 2 K. uk is the distance cost of transporting unit k, and ckij is the cost of transporting k on arc(i, j), and ckij = uk δij . Let F be the cost for installing a crossed turntable per period. The following notations are used: area of the machining zone; Ai H1 height of the upper row; height of the lower row; H2 L length of the AMHS loop; a fixed distance across the center spine; flow of commodity k on the directed arc; xijk yij DðkÞ PðkÞ S S

crossed turntable on arc (i, j); drop-off point of commodity k; pick-up point of commodity k; set of directed shortcut arcs, S  E; set of undirected shortcut arcs, S  S.

The width of the unit rectangle in the upper row is assumed to be one. Then, the width of the unit rectangle in the lower row is b = H1 =H2 . Note that L, Ai , a and b are used to calculate the arc length δij . The network flow model of the integrated shortcut design problem can be expressed as Minimize Z=

X X k2K ði, jÞ2E

ckij xijk +

X

Fyij

(3:1)

ði, jÞ2S

Subjected to X j2V

xijk −

X j2V

8 < − 1, if i = DðkÞ xkij = 1, if i = PðkÞ, ∀i 2 V and ∀k 2 K : 0, if otherwise

xkij , xjik ≤ yij , ∀ði, jÞ 2 S and ∀k 2 K

(3:2)

(3:3)

41

3.1 Modeling Methods Based on the Network Flow

xkij ≥ 0, ∀ði, jÞ 2 E and ∀k 2 K

(3:4)

yij 2 f0, 1g, ∀ði, jÞ 2 S

(3:5)

Assume that the flow sequence is given. Equation (3.1) represents the optimization objective function, which minimizes the reduction in handling costs while minimizing the increase in the shortcut costs. Equation (3.2) shows the flow conservation constraint for each commodity k. If (i, j) is not in the design, then the flow of each commodity k on the arc is zero in both directions. This is a multi-commodity network flow model with fixed expenditure variables on some arcs. Finally, this integrated design problem is solved by using iterative descent heuristic pairwise exchange heuristics, as shown in Figure 3.6, where N is the total number of workstations covered in the layout. The experimental results show that the integrated design method is superior to the traditional improved quadratic ensemble method. The modeling method of a wafer manufacturing system based on the network flow is often developed on the basis of some assumptions such as steady state and

set MIN = Integrated design cost based on the current flow sequence

for m = 1 to N – 1 and n = m + 1 to N Interchange departments m and n in the flow sequence and then update the layout design No

No

Is current layout feasible ?

Yes

Yes Update the integrated design cost Z by solving formulation (1)–(5)

Z < MIN ?

No

Yes Update the integrated design G = (V, E ); set MIN = Z Figure 3.6: Integrated design procedure.

m = N – 1 and n=N?

STOP G = (V, E ) and corresponding flow sequence is the final integrated design

42

3 Modeling methods of automated material handling systems in SWFSs

independence. The stochastic characteristics of the material handling process are simplified during modeling. And these conditions in reality may not be satisfied, and thus cannot guarantee the accuracy of its analysis results. Moreover, this method only analyzes the dynamic process of the system, and it is difficult to analyze the operation of the material handling system in the stochastic dynamic environment. The above limitation leads to the application limitation of this method.

3.2 Modeling Methods Based on the Queuing Theory The method based on the queuing network model usually describes the material handling system as a queuing network. The method is only suitable for the operation analysis of the small-scale material handling system. For large-scale, random and complex wafer manufacturing problems, the difficulty of solving problems will increase exponentially.

3.2.1 Basic Theory of the Queuing Theory Model Queuing theory [3–6] is also known as stochastic service system theory. First, the arrival time and the service time of the service object are statistically studied, and some statistical indexes (such as waiting time, queue length, busy period and so on) are obtained. Then, according to these indexes, the structure of the service system is improved or the service objects are reorganized. Hence, the service system can meet the needs of clients and optimize the cost or certain indicators. It is a branch of mathematical operations research disciplines. It also studies the randomization of queuing phenomena in service systems. It is widely applied in the random service systems with sharing resources such as computer networks, production and transportation, inventory and so on. Three aspects such as statistical inference, system behavior and system optimization are mainly investigated in the queuing theory. Its purpose is to correctly design and effectively run various service systems. Although the contents of the queue problems are different, they have the following common characteristics: 1. A person or thing that requests a service (e.g., a waiting patient, an aircraft requesting a landing, etc.) is regarded as a “customer.” 2. The person or thing that serves the customer (e.g., doctors, aircraft runways, etc.) is called “waiter.” The service system is formed by customers and waiters. 3. The customer randomly comes to the service system one by one (or a batch). The service time of a customer is not necessarily determinate. This randomness of the service process will result in a long queue of customers, and sometimes the waiter is free.

3.2 Modeling Methods Based on the Queuing Theory

43

In order to describe a given queuing system, the following components of the system must be defined: 1. The input process is the probability distribution that the customer comes to the service desk. On the basis of the original data, the experience distribution of the queuing problem is developed according to the rules of the customer arrivals. And then its theoretical distribution is determined by using the statistical method (such as chi-square test), and its parameter values are estimated. The probability distribution of the customer arriving at the service desk is a Poisson distribution, and the arrival of the customer is an independent smooth input process. “Smooth” means that the expected value of the distribution and variance parameters are not affected by time. 2. The queuing rules are those that define how customers queue and wait. The queuing rules generally consist of real-time systems and waiting systems. The rules of the real-time system are the rules that the customers leave when the service desk is occupied; while the rules of the waiting system are rules that the customers queue to wait for the service when the service desk is occupied. Waiting service rules include first-come first-served (FCFS) rule, random service rule, priority service rule and so on. The FCFS rule is investigated in this book. 3. The service agency consists of no waiter, one waiter or more waiters. The service agency provides service for an individual customer or for a group of customers. In the input process, most service times are random. It is always assumed that the distribution of service time is smooth. If ξn is the time required to service the nth customer, the probability distribution of the sequence of the service time {ξn, n = 1, 2, . . . } expresses the service mechanism of the queuing system. The successive service times ξ1, ξ2, . . . are independent, and the time interval sequence {Tn} for the arrival of any two customers is also independent. The purpose of studying the queuing problem is to study the operating efficiency of the queuing system, to evaluate the service quality, to determine the optimal value of the system parameters, to determine whether the system structure is reasonable and to design improvement measures. Therefore, it is necessary to determine basic quantitative indicators for the system. 1. Length: it is the number of customers (including waiting customers and served customers) in the system, and its expected value is Ls. The queuing length is the number of waiting customers in the queue, and its expected value is Lq. The number of customers in the system (length) = the number of customers waiting for service + the number of customers being served. Hence, the greater the Lq or Ls, the lower the service efficiency. 2. Waiting time: it is the period from the arrival time of the customer to the time to accept the service, and its expected value is Wq. The staying time is the period from the arrival time of the customer to the time to complete the

44

3.

3 Modeling methods of automated material handling systems in SWFSs

service, which is the total time spent by the customer in the system, and its expected value is Ws. staying time = waiting time + service time. Busy period: it refers to the continuous busy time of the service desk, which is the time from the arrival of the customer to the idle service desk until the service desk becomes idle again. This is the most important indicator; it is directly related to the waiter’s work intensity. The idle period refers to the length of time for which the desk is kept idle. Obviously, the busy period and the idle period are alternately present in the queuing system.

In addition, the service desk utilization (that is the percentage of the busy time of the waiter in the total time) is an important indicator in queuing theory. The standard form of the queuing model is presented by X/Y/Z/A/B/C, where X is the distribution type of customer arrival time; Y is the distribution type of service time; Z is the number of waiters; A is the system capacity; B is the number of customers; C is the service rule. The waiting queue model of the FCFS rule is mainly composed of three parameters: X/Y/Z. For example, “M/M/1/k/∞/FCFS” means that the customer arrival time and the service time obey the negative exponential distribution. “M/M/c” means that the customer arrival time and the service time obey the negative exponential distribution, and it consists of c service stations. “M/G/1” means that the customer arrival time obeys the negative exponential distribution and the service time obeys the general random distribution; it has a single service station. The input process of the system is subject to Poisson distribution, the service time obeys the negative exponential distribution and the queuing system of a single service station has the following three situations: 1. standard type: M/M/1 (M/M/1/∞/∞); 2. limited system capacity type: M/M/1/N/∞; 3. finite customer source type: M/M/1/∞/m. For a queuing system of a single service station, parallel C of the service desk, with the single-service desk similar to the multi-service queue system, has the following three situations: 1. standard type: M/M/C (M/M/C/∞/∞);

3.2 Modeling Methods Based on the Queuing Theory

2. 3.

45

limited system capacity type: M/M/C/N/∞; limited customer source type: M/M/C/∞/m.

For scheduling problems in AMHSs, there are waiting queues in the wafer carts and transport vehicles; therefore, many scholars investigate queuing problems.

3.2.2 Queuing Theory Modeling Process for AMHSs From the point of view of the queuing system, the Automated Guided Vehicle (AGV) system can be regarded as a resource, and a transport request can be regarded as an arriving customer. The required number of servers (the size of the fleet) depends on the parameters related to the customer arrival (transport request) and the service time (AGV handling time). In Figure 3.7, the allocation wait time is the delay time before the transport request for the workpiece is assigned to the cart. In terms of the queuing theory, the mean and variance of vehicle handling time are used to estimate the assignment waiting time. The idle times of the vehicle under different assignment rules are estimated by the analysis model proposed by Koo and Jang (2002) [14]. Once the assignment waiting time and the idle time have been calculated, the expected value of the waiting time of a part can also be estimated, and this parameter will be directly used to estimate the size of the fleet. The assumptions are presented as follows: 1. The handling time between the loading station and the unloading station is unique and deterministic. 2. The average delivery request rate for different sites is known. However, the interval time between delivery requests is subject to a certain probability distribution.

Delivery request

Vehicle assignment

Assisment waiting time

Loading

Vehicle idle time

Loaded vehicle handling time

Vehicle handling time (service time)

Part waiting time

Figure 3.7: Part waiting time in a material handling system.

46

3 Modeling methods of automated material handling systems in SWFSs

3.

If there is no delivery request waiting for the vehicle, then the vehicle stops at the current site to wait for the next request. 4. A vehicle can only respond to one delivery request at a time. 5. If more than one idle vehicle responds to a delivery request, the vehicle is selected according to the assignment rules. If multiple delivery requests wait for a free vehicle, the vehicle will respond to the delivery request according to the FCFS rule. In view of the above assumptions, the queuing network theory is used to investigate the fleet size problem in the system. First, the mean and variance of the idle time and the loaded handling time of the vehicle are estimated, and then the expected waiting time of the part during the handling process is estimated. The model parameters are defined as follows: n the number of pick-up/drop-off locations; the delivery request rate from location i to location j; fij the vehicle travel time from location i to location j; tij m the number of vehicles; ρ vehicle utilization; Sðk, iÞ the set of all locations that are closer to location i than location k; lu the sum of the loading time and the unloading time; ! n P n P fij ; F the delivery request rate between all locations F = i=1 j=1

q(k, i) fsi f dk

the probability that an idle vehicle at location k is selected when a delivery request is issued at location i; the proportion of delivery requirements from location i; the proportion of the delivery requirements to location k.

The loaded vehicle handling time includes the loading time and the unloading time; its mean and variance are calculated as follows:

Eðtl Þ =

" n n  X X i=1

Vðtl Þ =

# (3:6)

j=1

" n n n X X i=1

ðfij =FÞðtij + luÞ



2

ðfij =FÞðtij + luÞ

o

# − E2 ðtl Þ

(3:7)

j=1

Its expected value is a weighted average of all possible values. Formulas (3.6) and (3.7) are independent to the selection rule of the idle vehicle.

47

3.2 Modeling Methods Based on the Queuing Theory

1.

The idle vehicle handling time Suppose that the idle time of the vehicle is independent. Therefore, the number of idle vehicles Z at the moment of issuing the delivery request follows a quadratic distribution: bðm, 1 − ρÞ, i. e., Pð zÞ = prðZ = zÞ = Czm ð1 − ρÞz ρm − z , where Czm =

m! ð m − zÞ!z!

Case 1: If there is no free vehicle response to the delivery request, the delivery request is added to the waiting queue. The probability of this situation is ρm or P (0). In this case, when the vehicle is idle, the delivery request is responded in accordance with the FCFS rule. Case 2: If at least one vehicle responses when a delivery request arrives, the free vehicle is arranged according to the assignment rule. The probability of this m P situation is ð1 − ρm Þ or PðzÞ. z=1

Then the idle handling time is estimated by using different dispatching rules. (1) Random vehicle selection

Eðte Þ =

n X

" fsi

n X

i=1

Vðte Þ =

n X

" fsi

ðfdk tki Þ

(3:8)

j=1

n X

i=1

where fdk =

#

# − E2 ðte Þ

2 fdk tki

(3:9)

j=1

n n 1X 1X fik , fsi = fij . F i=1 F j=1

(2) Longest idle vehicle selection and least utilized vehicle selection are the same as eqs (3.8) and (3.9). (3) Nearest vehicle selection

Eðte Þ =

n X z=1

" Pð z Þ

m X

" fsi

i=1

V ðte Þ =

n X

## fqðk, iÞtki g

+ Pð0Þ

i=1

n P

n X

" fsi

i=1

 Pð z Þ

z=1

+ Pð0Þ

m P i=1

n P i=1

n X

# ðfdk tki Þ

(3:10)

i=1

 n  P 2 fsi fqðk, iÞtki g i=1

  n P 2 fsi ðfdk tki Þ − E2 ðte Þ i=1

(3:11)

48

3 Modeling methods of automated material handling systems in SWFSs

" qðk, iÞ = 1 −

#

P

qðr, iÞ Aðk, iÞ

r2Sðk, iÞ

where Aðk, iÞ = 1 − 1 − P

!z

fdk

r ∉ Sðk,

iÞfdr

, i, k = 1, 2, ..., n

(4) Farthest vehicle selection is the same as eqs (3.10) and (3.11). Assume that the handling time and the idle time of a loaded vehicle are independent; then Eðtv Þ = Eðtl Þ + Eðte Þ

(3:12)

V ðtv Þ = V ðtl Þ + V ðte Þ

(3:13)

The queuing theory is used to estimate the waiting time of the assignment process. According to the G/G/m queuing system proposed by Kimura (1991) [15], the assignment waiting time is calculated by   Wq = Wo c2a + c2v gw=2

(3:14)

where 8 >
F0.05 ð1, 23Þ 4.76>F0.05 ð1, 23Þ 4.65>F0.05 ð1, 23Þ 3.34>F0.05 ð1, 23Þ

0.9 Comprehensive function D

0.8 0.7 0.6

MHAFLC

0.5

STD

0.4 CLAB

0.3

RLWT

0.2 0.1 0 Scenario-1

Scenario-2

Scenario-3 Scenario-4

Figure 3.15: Comparison of D-function values of proposed dispatching method and other rules.

Furthermore, it is found that the proposed MHAFLC dispatching method has better performance except for vehicle utilization in all scenarios. In terms of D-function value, MHA-based dispatching methods have the best performance in comparison with the other rules. However, the fluctuation of its D-function value indicates that the robustness needs to be strengthened. The proposed fuzzy-logic-based control method is applied to dynamically adjust the weight and therefore improve the robustness. Next, how to use the simulation model to evaluate different scheduling rules? In the one-way double closed-loop Interbay material transportation system, the eventdriven vehicle assignment strategy has three decision points. 1) When the wafer workpiece needs to be moved to the next storage depot, the transport loop needs to be selected. There are four rules for loop selection: Shortest Distance First (SD), Work in Progress (WN), Work Distance Rule (WD) and Work in Process Rule (WR) on robot transfer mechanism (RTM) 2) Allocate empty vehicles for wafer workpieces. For vehicle assignments, only the recent vehicle priority rule (NV) is available. 3) Arrange the wafer workpiece to be moved to the free vehicle. There are two rules: first encounter first service (FEFS) and the longest waiting time priority rule (LWT). According to the above three decision points and seven rules, you can design eight kinds of rule combinations, as shown in Figure 3.16.

66

3 Modeling methods of automated material handling systems in SWFSs

Cassette-initiated dispatching rules

Vehicle-initiated dispatching rules

Loop selection-vehicle assignment

Cassette selection

1. SD-NV 1. FEFS 2. WN-NV 3. WD-NV 2. LWT 4. WR-NV

Figure 3.16: Dispatching rules based on casette and vehicle initiation.

Interbay system Level 2 INRERBAY

STK

Loop

Level 1

Level 0

Arrival Crane Source Exit buffer

Bay

I/O RTM Vehicle

Userdefined objects Basic objects Figure 3.17: Hierarchy of an interbay simulation model.

In order to evaluate the performance of the eight rules as shown in Figure 3.16, a discrete time simulation model of an Interbay material handling system is established based on the simulation software. With the base objects provided by eM-Plant, users can customize more complex objects such as loops that contain buffers, tracks and trolleys. Figure 3.17 shows the objects and their hierarchical relationships of the simulation model for Interbay material handling systems.

3.6 Modeling Methods Based on the Petri Net

67

In general, productivity in the Interbay system has the highest weight. Two factors affect the performance of the assigned rules: 1) the flow rate of the wafer workpiece and 2) the number of vehicles. The wafer workpiece flow rate can be set to 345,690 or 1,035 cassette h-1. The number of vehicles can be set to (per loop) 8, 12 or 16. Therefore, a total of nine kinds of scenes are possible. Statistical tests using ANOVA show that the number of repetitions of the simulation model does not significantly affect the statistical data of the simulation model. Therefore, each simulation model is repeated only three times. The total number of simulation experiments is 8 (rule) × 9 (scene) × 3 (repeat) = 216 times. Experimental results show that the assignment rule has significant influence on average transit time, productivity, waiting time and vehicle utilization. Among them, the combination of SD-NV and FEFS rules is obviously superior to other combination of rules. In addition, through the analysis of the corresponding curved surface, the optimal combination of the flow rate and the number of vehicle is obtained. In general, the scheduling model based on the simulation model can describe the AMHS scheduling process. However, the method also has serious disadvantages such as long modeling period, low modeling efficiency and poor reusability.

3.6 Modeling Methods Based on the Petri Net The Petri net is a tool developed by C.A. Petri in his doctoral thesis in 1982 as an asynchronous concurrent operation to describe the elements of the system. After several decades of development, the Petri net theory has been widely used in many fields such as computer, electronics, machinery, chemistry and physics, and has become a powerful tool for modeling AMHS [10].

3.6.1 Basic Theory of the Petri Net Model Petri nets are supported by formalized mathematical methods and can be used to express the static structure and dynamic changes of the discrete event dynamic system (DEDS). It is a structured DEDS description tool that describes the logical relationships of asynchronous, synchronous and parallel systems. These descriptions are used to check and analyze the performance of the system. For example, the Petri net is adopted to describe machine utilization, productivity, reliability and other indicators. A Petri net is a triplet: N = ðP, T, FÞ where P = fp1 , p2 , ..., pm g is the set of places; T = ft1 , t2 , ..., tn g is the set of transitions;

68

3 Modeling methods of automated material handling systems in SWFSs

F = ðP × TÞ ∪ ðT × PÞ is the input function and the output function set, which is regarded as the flow relationship. A triplet N = ðP, T, FÞ is formed to be a net structure under the practical constraints. The necessary and sufficient conditions for the triple network to form a net are as follows: ① P ∩ T = ϕ defines that the places and the transitions are two different kinds of elements; ② P ∪ T ≠ ϕ indicates that there is at least one element in the net; ③ F = ðP × TÞ ∪ ðT × PÞ establishes the unilateral links from the places to the transitions and vice versa, and it is not possible to provide the direct link between similar elements. The state of a process is modeled using tokens in places, and the transitions of states are modeled by using transitions. A token represents the state of a thing (person, goods, machine), information, condition or object; a place represents a location, a channel or a geographic location; a transition represents an event, transformation or transmission. A process consists of a current state, reachable state, unreachable state and termination status. The Petri net is an effective tool to investigate discrete manufacturing systems. It has the dual functions of graphical representation and mathematical description in dealing with dynamic discrete events and complex systems. It is especially suitable for modeling the concurrent behavior, and it is widely used in modeling AMHSs, modeling wafer processing systems and evaluating scheduling performances.

3.6.2 Petri Net Modeling Process for AMHSs An agent can make decisions and control based on common goals and information. In terms of this advantage of an agent, the scheduling performance evaluation of the wafer processing system is modeled using an agent-oriented colored timed Petri net (AOCTPN) developed by Zhang Jie and Zhai Wenbin [11–13]. The modeling process is presented as follows. In order to reduce the complexity and increase the reusability of re-entrant system model, the agents of re-entrant system with similar structure and function are aggregated to form the layered structure, which includes basic agent layer, unit layer and system layer (as shown in Figure 3.18). 1. Basic agent layer modeling As the basic element of a re-entrant system model, the agents in the basic agent layer include physical agents and the logical ones. Corresponding to the manufacturing resources of a re-entrant system, the physical agents include machine

3.6 Modeling Methods Based on the Petri Net

Reentrant manufacturing system model

System layer

Unit layer

Basic agent layer

69

One-piece equipment group model

Agent model

..

Batch equipment group model

Agent model

Agent model

..

Agent model

Other equipment agent model

Buffer agent model

Feed agent model

Routing agent model

Figure 3.18: Hierarchical structure of logistics network model.

MReady_p

Failure3_ts

Repair_ts

MRepair_p

Failure1_ts Setup_ti

Setup_p

IMi1

OMi1 Failure2_te

Processing_p OMi2

Enter_ti Setup_te

OK_te

Figure 3.19: AOCTPN based machine agent.

agents, buffer agents and so on. The logical agents with dispatch or control functions include lot release agents, dispatching agents and so on. The machine agent is the core in the machine layer, whose model characterizes the interior behavior of the agent, such as machine setup, operation or rework, failure and maintenance, and so on. According to the part types, the machines are classified into single-lot processing machines and batch processing machines. The reasoning process of the single-lot-machine-agent model is shown in Figure 3.19. The token in the input message place IMi1 represents the

70

3 Modeling methods of automated material handling systems in SWFSs

Table 3.8: Color data type. Data interpretation

Elements

Type name

LI P S

Lot NO Product ID Lot stage

PT

Product processing time

ST KMG m E L

Key machine setup time Key machine group NO Key machine NO Key machine state Result of lot processing

Natural numbers {p1、 . . . 、pj、 . . . 、pm} {S1、NS1、 . . . 、Sj、NSj、 . . . 、Sn、NSn} {PT1、NPT1、 . . . 、PTj、 NPTj、 . . . 、PTn、NPTn} {ST1、 . . . 、STj、 . . . 、STn} Natural numbers Natural numbers {eok、er、ef} {lok、lr}

Table 3.9: Color sets of the single lot processing machine agent. Name

Function interpretation

Color set

IMi1 MReady_p Enter_ti Setup_p Setup_ti Setup_te Processing_p

Input message of lot waiting for processing Machine being idle Transfer to machine Check machine setup Process without setup Machine setup Start lot processing

Failure1_ts Failure2_te Failure3_ts OK_te Mrepair_p Repair_ts OMi1 OMi2

Machine failure and lot rework Machine idle and lot rework Machine failure Lot processing Start machine repair Machine repair Output message of rework Output message of next process stage

LI × {Pi} × {Sj} × KM KM × {eok } × {Pi} × {Sj} LI × {Pi} × {Sj} × KM × {eok } LI × {Pi} × {Sj} × KM × {eok } {LPk } × {LSm} × KM × {eok } {Pi} × {Sj} × KM × {eok } × {STj} LI × {lok } × {Pi} × {Sj} × KM × {eok } ∪ LI × {lr} × {Pi} × {Sj} × KM × {eok } ∪ LI × {lr} × {Pi} × {Sj} × KM × {ef} LI × {Pi} × {Sj} × KM × {ef } LI × {Pi} × {Sj} × KM × {eok } × {PTj} KM × {ef} × {Pi} × {Sj} LI × {Pi} × {Sj+1 } × KM × {eok } × {PTj} KM × {er} × {Pi} × {Sj} KM × {er} × {Pi} × {Sj} LI × {Pi} × {Sj} LI × {Pi} × {Sj+1 }

lot-processing request messages sent by the dispatch agent. The data type of different colors is shown in Table 3.8. The color sets of the single-lot processing machine agent are shown in Table 3.9. The token in the place MReady_p corresponds to machine idleness. When the machine is idle and receives the lotprocess-request message, the immediate transition Enter_ti is fired, and the lot token enters the place Setup_p. The color of the immediate transition Enter_ti is

3.6 Modeling Methods Based on the Petri Net

71

{LPk} × {LSm} × KM, which records the last process step of the lot. The deterministically timed transition Setup_te represents the setup time of the machine. If the color {Pi} × {Sj} × KM of the token in the place Setup_p is the same as the color of the transition Setup_ti, then the transition Setup_ti is fired and the lot token directly enters place Processing_p for processing. Otherwise, the transition Setup_te is fired and the machine starts to set up. The place Processing_p may fire one of three kinds of the stochastic transitions, Failure1_ts, Failure2_te or OK_ts, which respectively correspond to machine failure, lot fault and finish. According to the rework rate of a lot, whether the lot reworks or not is decided. If the lot does not rework, then the immediate transition OK_te is fired. After a period of process time PTj, the machine becomes idle and the machine token enters MReady_p, and the lot token in place Processing_p enters the output message place OMi2. If the lot needs rework, then machine failure is decided by the mean time to failure rate (MTTF). In the case of the machine failure, the transition Failure1_ts is fired and the machine starts repair. The machine token enters place Mrepair_p. At the same time, the lot token enters the output message place OMi1 to send a rework request. If the machine is okay, the transition Failure2_te is fired and the machine is idle. The machine token enters place MReady_p. Similarly, the lot token enters output message place OMi1 to send rework request. Since the process time is shorter than the failure cycle of the machine, it supposes that the machine never fails during the processing period. As the machine is idle, the failure of the machine is decided by MTTF. If the machine fails, the transition Failure3_ts is fired. The time for machine repair is given by the mean time to repair (MTTR) rate. The case of the batch processing machine agent is complicated, and its figure is the same as Figure 3.19. The color set of IMi1, Enter_ti, Setup_p, Setup_ti, Setup_te, Failure1_ts, Failure2_te, OK_ts, OMi1 and OMi2 depends on the batch size of the machine. For example, if the batch size is two, the color set of Enter_ti is LI1 × {Pi} × {Sj} × LI2 × {Pm} × {Sn} × KM. The other agent models based on AOCTPN are shown in Figure 3.20. In the actual production, it is assumed that the capacity of the non-bottleneck equipment in the buffer zone is infinite. In the non-critical equipment agent model, the machining process is simplified to a fixed time delay, and the time is related to the total processing time of non-critical processes. The agent model invokes the interactive protocol network model through message bits OMi2 and IMi2. Through the multi-agent negotiation, the scheduling process of the processing tasks is completed.

2.

Unit layer modeling The unit layer agent belongs to the composite agent and consists of several processing machine agents and machine group agents with the same machining

72

3 Modeling methods of automated material handling systems in SWFSs

Processing_p

Setup_p

Enter_ti

Exit_ti

Enter_ti

(a)

Exit_ti

Ok_te

(b) OMi2

Send_ti

IMi2

Receive_ti

Release_p

OMi1

OMi1

IMi1

Enter_ti

(c)

OMi1

OMi1 IMi1

IMi1

Release_ti

Dispatch_ti Receive_p Exit_ti Send_p

(d)

Figure 3.20: Other agent models based on AOCTPN: (a) the buffer agent model; (b) the non-key equip agent model; (c) the dispatch agent model; and (d) the lot release agent model.

function. Each basic agent is connected to form a complete AOCTPN model according to specific control logic; that is, the communication transition determines the cooperative and interactive relationship between the manufacturing resources of the unit layer. In the specific unit layer model, the triggering mode and the triggering order of the communication transition determine the control logic of the unit layer. The triggering processes of some traffic changes can be determined by the characteristics of the Petri net (Figure 3.21(a)). However, the triggering processes of other communication changes should introduce the corresponding multi-agent interaction protocol in the model, which is regarded as the conflict when the characteristics of Petri nets cannot be determined and multi-agent cooperative decision is needed. Its corresponding multi-agent cooperative strategy is regarded as the multi-agent interaction protocol of conflict discrimination. When a collision is detected, the device agent issues a conflict resolution request to the machine group agent. The machine group agent calls the multi-agent interaction protocol to complete the conflict resolution. The conflict in the model can be classified into three classes according to the specific situation: 1. There is a conflict between the communication changes; that is, two or more communication changes have the same input bit. Therefore, it is necessary to determine which communication transition is triggered by the input bit. This is consistent with the conflict in the general Petri net. This kind of conflict usually occurs when a part in a unit chooses a processing machine, and the conflict can be solved by a multi-agent interaction protocol (Figure 3.21(b)).

3.6 Modeling Methods Based on the Petri Net

IMi1

73

IMi1 Ti1

Ti1

Tk1 OMj1

OMj1 (a)

OMk1 (b) IMj1

IMi1

IMi1 Ti1 Ti1

(c)

Ti2 OMk1 (d)

Figure 3.21: Conflicts in the AOCTPN model.

2.

3.

In collisions between tokens in bits that trigger communication transitions, where the number of tokens in the input bits of the communication transition is greater than the number of tokens required to trigger the transition, it is necessary to determine which token to trigger for the corresponding transition. This conflict usually occurs when multiple parts in the buffer are waiting to be processed, and the conflict can be resolved according to the multi-agent interaction protocol (Figure 3.21(c)). Conflicts between bits trigger communication transitions, in which the input bit IMk1 can be triggered by a different output bit IMi1 or IMj1 (Figure 3.21(d)). The solution at this time is to determine which communication transition triggers first. This kind of conflict often occurs when parts which are processed by different machines compete with the processing machine immediately after the process, and the multi-agent interaction protocol can be used to solve the conflict. The reasoning process of the key machine group agent is shown in Figure 3.22. The token in the input message place IMj1 corresponds to the lot-processing request messages sent by the buffer agent. By firing the immediate transition ID_ti, the lot token enters the dispatch agent place Dispatch_ap. Guided by the coordination mechanism, the dispatch agent optimally finds the lot token that should enter the machine agent place KM_ap_i and the communication transition DK_ti_i that should be fired. If a lot needs rework, the communication transition KD_ti_i is

74

3 Modeling methods of automated material handling systems in SWFSs

Table 3.10: Color sets of the single lot processing machine group. Name

Function interpretation

Color set

IMj1

Input message of lot waiting for processing Request of dispatch Select processing lot and machine Request of processing Lot processing Request of output Message of lot rework Output message of next process stage

LI × {Pi} × {Sj}

ID_ti Dispatch_ap DK_ti_i KM_ap_i KM_ti_i KD_ti_i OMj1

LI × {Pi} × {Sj} LI × {Pi} × {Sj} × KM LI × {Pi} LI × {Pi} LI × {Pi} LI × {Pi} LI × {Pi}

KM _ap_1

KM_ti_1

KM _ap_i

KM_ti_i

DK_ti_1

× {Sj} × KM × {Sj} × KM × {PTj} × {Sj+1} × {Sj} × {Sj+1}

KD_ti_1

IMj1

Dispatch _ap DK_ti_i ID _ti

OMj1

KD_ti_i

DK_ti_m

KM _ap_m KM_ti_m

KD_ti_m

Figure 3.22: AOCTPN based machine group.

fired. Otherwise, the communication transition KM_ti_i is fired and the lot token enters the output message place OMj1. The color sets of the single-lot processing machine group agent are shown in Table 3.10. In the case of the batch-processing machine agent, the color sets of its place and transitions depend on the batch size. 3.

System layer modeling During the lot processing, the main waiting time happens in the bottleneck machine group. Hence, the bottleneck machine group is called the key

3.6 Modeling Methods Based on the Petri Net

75

machine group. The system model focuses on the key machine group while the non-key machine group is treated as infinite possessing ability and is simplified into deterministic time delay. Since the buffer and material handling system capacity is bigger than the work-in-process (WIP) requirement, its constraints are neglected. The re-entrant system model in the system layer is shown in Figure 3.23. The release agent place Release_ap sends a message to the agent place of the material handling system. After lots transfer, the lot token enters the buffer agent place Buffer_ap_i. When the agent place of the key machine group KMG_ap_i receives the processing request from the buffer agent place Buffer_ap_i, the lot token enters KMG_ap_i. After the key processing step KMG_ap_i is completed and the non-key processing step NKMG_ap_i is completed, the lot token enters check-out agent place Check-out_ap. If the lot has finished all steps, it enters exit agent place End_ap. Otherwise, the lot token reenters the Routing_ap for the next process step. The color sets of the system layer are shown in Table 3.11. The AOCTPN method can describe autonomous coordination scheduling and control logic in material handling systems, but it cannot describe information and knowledge required for scheduling decision and running process. Therefore, it is difficult to effectively integrate the material handling scheduling process with the performance evaluation model.

Table 3.11: Color sets of the system layer. Name

Function interpretation

Color set

Release_ap RR_ti Routing_ap RB_ti_i Buffer_ap_i BK_ti_i

Lot release Request of lot transfer lot transfer Request of entering buffer Enter buffer Request of lot processing by key machine group agent Key machine group processing

LI × {Pi} × {Sj} LI × {Pi} × {Sj} LI × {Pi} × {Sj} × KMG LI × {Pi} × {Sj} × KMG LI × {Pi} × {Sj} × KMG LI × {Pi} × {Sj} × KMG

KMG_ap_i KN_ti_i NKMG_ap_i

Request of lot processing by non-key machine group agent Non-key machine group processing

NC_ti_i Check-out_ap CR_ti CE_ti End_ap

Request of checking exit Checking whether lot exit Send message of processing next stage Send message of lot exit Lot exit

LI × {Pi} × {Sj} × KMG × KM × {PTj} LI × {Pi} × {Sj+1} × {NSj} LI × {Pi} × {Sj+1} × {NSj} × {NPTj} LI × {Pi} × {Sj+1} × {NSj+1} LI × {Pi} × {Sj+1} × {NSj+1} LI × {Pi} × {Sj+1} × {NSj+1} LI × {Pi} × {Send} × {NSend} LI × {Pi} × {Send} × {NSend}

76

3 Modeling methods of automated material handling systems in SWFSs

CR_ti

Buffer_ap_1 Release_ap

RB_ti_1

RR_ti

Routing _ap

BK_ti_1

Buffer_ap_i

RB_ti_i

NKMG_ap_1

KN_ti_1

KMG_ap_i

BK_ti_i

Buffer_ap_n

RB_ti_n

KMG_ap_1

BK_ti_n

NKMG_ap_i

KN_ti_i

KMG_ap_n

NC_ti_1

NC_ti_i

Check-out_ap

CE_ti

NKMG_ap_n

KN_ti_n

NC_ti_n

End_ap

Figure 3.23: AOCTPN based system layer model.

3.7 Conclusion This chapter presents several modeling methods of AMHSs regarding two aspects: the basic theory and its applications in scheduling problems. The modeling methods such as network flow model, queuing theory model, mathematical programming model and Markov model are often developed on the basis of the steady state. In the modeling process, the stochastic characteristics of the material handling process are simplified, these conditions may not be satisfied in practice, and the accuracy of the analysis results cannot be guaranteed. It is difficult to obtain the solutions for large-scale problems. Both the simulation model and the Petri net model are effective tools to study the discrete manufacturing system, which can describe and analyze the operation process of an AMHS. In particular, Petri nets have the dual functions of graphical representation and mathematical description in dealing with the characteristics of dynamic discrete events and complex systems, especially for modeling concurrent behavior, and are particularly useful for modeling concurrency behavior as a powerful tool to evaluate the scheduling performances. Different modeling methods can solve actual problems with actual demands. For example, the queuing theory and the Markov model can be used to achieve better results at lower cost to analyze the statistical properties of an AMHS. The simulation model or the Petri net model can be used to clearly describe the operations of an AMHS.

References

77

References [1] [2] [3] [4] [5] [6]

[7]

[8] [9] [10] [11] [12]

[13]

[14]

[15]

Jiang, Q.Y., Xie, J.X., Ye J. Mathematical Model. Beijing: Higher Education Press, 2003. Chen, B.L. Theory and Algorithms of Optimization. Beijing: Tsinghua University Press, 2005. Averill, M. Law. Simulation Modeling and Analysis (Fourth Edition). McGraw Hill Higher Education, New York, 2007. Fishman, G.S. Statistical analysis for queuing simulations. Management Science, 1999, 20(3). Curry, G.L., Peters, B.A., Lee, M. Queuing network model for a class of material handling systems. International Journal of Production Research, 2003, 41: 3901–3920. Roeder, T., Govind, N., Schruben, L. A queuing network approximation of semiconductor automated material handling systems: how much information do we really need?. Proceedings of the 2004 Winter Simulation Conference. 2004, pp. 1956–1961. Nazzal, D., Mcginnis, L.F. Queuing models of vehicle-based automated material handling systems in semiconductor Fabs. Proceedings of the 2005 Winter Simulation Conference. 2005, pp. 2464–2471. Johnson, M.E., Brandeau, M.L. An analytic model for design and analysis of single vehicle asynchronous material handling systems. Transportation Science, 1994, 28: 337–353. Kobza, J.E., Yu-Cheng, S., Reasor, R.J. A stochastic model of empty-vehicle travel time and load request service time in light-traffic material handling systems. IIE Transactions, 1998, 30: 133–142. Peterson, J.L. Petri Net Theory and the Modeling of Systems. Prentice-Hall, New Jersey, USA, 1981. Zhang, J., Zhai, W., Yan, J. Multi-agent-based modeling for re-entrant manufacturing system. International Journal of Production Research, 2007, 45(13): 3017–3036. Zhai, W.B., Chu, X.N., Zhang, J., Ma, D.Z. Research on AOCTPN based modeling technology of semiconductor fabrication line. Computer Integrated Manufacturing Systems, 2005, 11(3): 326–329. Zhai, W.B., Zhang, J., Yan, J.Q., Ma, D.Z. Agent -oriented colored petri-net based interactive protocol modeling technologies of semiconductor fabrication line. Journal of Shanghai Jiaotong University, 2005, 39(7): 1150–1154. BS Koo, HS Jang. Evaluation on the impact of Lowest Bid Contracts on Site Operations in times of Severe Economic Downturn. Korean Journal of Construction Engineering and Management, 2009, 10 (6): 146–153. Kimura, Approximating the Mean Waiting Time in the GI/G/s Queue. Journal of the Operational Research Society, 1991, 42(11): 959–970.

4 Analysis of automated material handling systems in SWFSs There have been many reports of investigation on the performance of analytical modeling approach for automated material handling systems. The most common and practical approach is the simulation model. Although the use of the simulation tool is advantageous in that all details can be taken into account, it is time-consuming and error prone to generate statistically reliable results. Especially when the dynamic system changes, e.g., the adjustment of the layout, the increase of machine or vehicles and the setup of some operational rules, the reconstruction of the simulation model is tedious. To avoid the shortcomings of simulation approaches, various mathematical models which are more efficient have been proposed to evaluate the performance of the automated material handling system (AMHS), such as queuing model, queuing network model and Markov chain model (MCM). Compared with the simulation tools, these analytical models may need to be improved in terms of the effectiveness (i.e., the accuracy); however, their runtime efficiency is much higher. This is critical in some environment with high demand for real-time decision; for example, the model is used to test various schedules. The mathematical models could evaluate the performance of schedules in several minutes while the simulation model will run a couple of hours to evaluate, which normally is not acceptable. This chapter proposes a modified Markov chain model (MMCM) for more general AMHSs, in which the distinctive AMHS’s features including shortcut rail, vehicle blockage and multi-vehicle transportation are taken into account. On the basis of the proposed model, the performance of the AMHS is efficiently analyzed and evaluated.

4.1 Running Process Description In semiconductor manufacturing, multiple layers of miniature circuitry are built upon a wafer. For each layer, a sequence of similar processes must be undertaken repeatedly, often on the same pieces of equipment. Between two processing steps, an AMHS is used to transport wafer lots. The layout of an AMHS is similar to that of a flow shop with functional areas, and can be separated into Interbay and Intrabay AMHS (as shown in Figure 2.2). Interbay and Intrabay AMHS are connected by stockers [1]. In a typical AMHS, the distance between adjacent stockers is 20 m. The distance between the pick-up port and the drop-off port of each stocker is 5 m. And the total Interbay rail loop is 320 m long. Since the configuration and constitution of Interbay AMHS and Intrabay AMHS are similar, the Interbay AMHS is investigated in this chapter. The Interbay AMHS typically consists of transportation rails, stockers, automated vehicles, shortcut and turntable. The transportation rails are usually monorail https://doi.org/10.1515/9783110487473-004

4.1 Running Process Description

79

systems, where automated vehicles transport wafer lots along the rails between different stockers. Each stocker has an input/output port and provides temporary storage for work in process of wafer lots. Each stocker is connected with an Intrabay AMHS. Automated vehicles are used to carry and transport wafer lots. Shortcuts are special transportation rails that are used to connect two transportation rails with opposite directions. Shortcut can shorten the vehicle’s transportation distance between stockers significantly. Turntables are the intersection of shortcut and transportation rails. Due to the space restriction in the Interbay AMHS, vehicle’s movement is along a unidirectional closed loop and vehicles cannot pass each other, even when a vehicle stops to drop-off or pick-up a wafer lot at a stocker. Let LðmÞ refer to the Interbay AMHS with m vehicles. Denote S as the set of stockers in the Interbay system, S = fsi , i = 1, ..., ng. si is the ith stocker in the Interbay system. Each stocker has a pick-up station and a drop-off station where wafer lots are dropped off and picked up by vehicles, respectively. Let spi and sdi denote the pick-up station and the drop-off station. Then, an Interbay AMHS with n stockers consists of 2n stations. As the pickup and drop-off stations are represented as nodes, and transportation rails and shortcuts are represented as directed arcs, the network representation of the Interbay AMHS can be illustrated as shown in Figure 4.1. The continuous material handling process is defined as the vehicle’s discrete travel process at the pick-up/drop-off station to describe material handling logic of the Interbay AMHS and analyze its performance indices effectively. In the Interbay AMHS, the vehicles constantly move on the unidirectional loop and transport wafer lots based on first come first service (FCFS) rule. Without loss of generality, given that an empty vehicle approaches stocker si , it passes through the drop-off station sdh and travels to the pick-up station sph . If there is a wafer lot waiting at sph , then the vehicle

stops and picks it up. After a constant loading time tl , the vehicle moves to the destination stocker according to the transportation path ph = ðsph , sdh + 1 , sph + 1 , ..., sdn , ...Þ (given the destination position of the current lot is stocker sn , sn 2 S). The vehicle

sd1

s1p

snp

snd

...

...

sdi

sip

sdi+1

spi+1

...

sdh

shp

p si+k

sdi+k

p si+k–1

sdi+k–1

...

sph+1

sdh+1

Vehicle

Drop-off and pick-up station

Figure 4.1: Network representation of the Interbay AMHS.

80

4 Analysis of automated material handling systems in SWFSs

does not stop at stations sdh + 1 , sph + 1 , ..., and spn − 1 unless it is blocked by other vehicles at downstream stations. After vehicle arrives at stocker sdn and drops off its load, it travels to spn and inspects if there are wafer lots waiting at spn , and so forth until it encounters a waiting wafer lot.

4.2 Methods for Analyzing Running Characteristics On the basis of the AMHS model, the internal changeover process and operation mechanism of an AMHS are analyzed. The static and dynamic factors that affect the performance of an AMHS are determined, and the changing rules and characteristics of an AMHS are revealed, which provide a foundation to optimize the AMHS. As mentioned in Chapter 3, modeling is the process to abstract the system. With different mathematical or computing tools, we can establish different system models from different angles. These different models have different characteristics and application fields, while the corresponding analytical methods depend on the established system models. For example, analysis methods based on queuing network model, Markov model and perturbation analysis mainly explain the running process of an AMHS in the statistical level. Analytical methods based on formal languages and automata, Markov chains and Petri nets focus on analyzing a system’s internal behavior in the logical level. Analytical methods based on a finite recursion process and a minimax algebra model study system’s trajectory and characteristics in the time level. Since several common modeling methods in a wafer manufacturing system have been introduced in the previous chapter, this section mainly presents the advantages and disadvantages of the analytical methods. The analytical method based on the network flow model generally establishes the running analysis model of an AMHS by simplifying the AMHS into a deterministic directed graph model, and regarding the material handling time as the deterministic time. Based on the model, we can study the performance of an Interbay system and the related influence factors including vehicle idle time, vehicle quantity and vehicle speed. However, this method relies on assumptions such as steady state, independence and so on. In the process of modeling, random characteristics of the material handling process are simplified, which cannot be satisfied in practice, and thus cannot guarantee the results of the analysis accuracy. Therefore, the network flow model is suitable for the statistical analysis of a stationary system, and it is difficult to describe the stochastic characteristics of the running process of an AMHS. The analytical methods based on the queuing network model generally models the AMHS as a queuing network, and depicts different running states with Markov model. Then the model can be used to analyze the impacts of handling shortcuts location on quantity of WIP, the relationship between the speed and load of vehicle

4.3 Modified Markov Chain Model

81

and system throughput, the impacts of vehicle number on system throughput and vehicle average transportation time, the impacts of vehicle speed and turntable system rotation speed on system’s performance, the impacts of vehicle number on material waiting time, the impacts of vehicle blockage on regional control AMHS performance and so on. The queuing network model can effectively describe the random behavior of an AMHS. However, it is difficult to model the characteristics of multiple vehicles blockage and empty vehicles. Therefore, this method only applies to analyze small-scale AMHSs, while for large-scale and complex AMHSs it is easy to cause a massive explosion of model states. The analytical method based on the mathematical programming model analyzes its running characteristics and optimizes the system’s operation parameters. The method is mainly used to analyze the effects of handling shortcuts and track positions on vehicle handling time, the effects of vehicle blocking time, vehicle speed and vehicle number on vehicle empty time and vehicle load rate, and the effects of vehicle number on AMHS’s throughput and production cost. Nevertheless, it is difficult to apply the method in a stochastic and large-scale environment. The analytical method based on the Markov model considers the Markov properties of state change sequences in vehicle handling, loading and unloading wafers. The method simplifies the AMHS scheduling problem into the mathematical model by defining vehicle-related parameters, storage-related parameters and states-related parameters, and establishes the running characteristic analysis model of the AMHS. It can be used to analyze the impacts of a single vehicle’s speed and load on AMHS performance, the vehicle’s performance under different running speeds and loading/ unloading times, and the impacts of wafer’s waiting time and vehicle number on AMHS performance. However, the Markov model can easily cause the state space explosion of the model in a large-scale environment, and it is very difficult to solve the model. To summarize, due to the complexity of the AMHS, it is difficult to apply the network flow model, the queuing theory and the mathematical programming model to the system for operation analysis. This chapter presents an improved Markov chain model for performance analysis of the AMHS. The method considering characteristics of shortcut track’s layout can accurately express operation characteristics of an AMHS so as to analyze the performance of the complex AMHS [2].

4.3 Modified Markov Chain Model So far, many approaches for AMHS performance analyzing have been proposed, such as queuing theory model, queuing network model, Markov chain model (MCM) and so on. Due to the complexity issue, few literature can be found upon modeling for the AMHS with shortcuts and blocking. In this section, an MMCM for more general

82

4 Analysis of automated material handling systems in SWFSs

AMHSs is proposed, in which the distinctive AMHS’s features including shortcut rail, vehicle blockage and multi-vehicle transportation are taken into account. On the basis of the proposed model, the performance of the AMHS is effectively analyzed and evaluated.

4.3.1 Notation and System Assumption 1.

2.

Parameters related to stockers S – Set of stockers in the system; n – Quantity of stockers in the system; si – The ith stocker in the system, si 2 S, i 2 f1, 2, ..., jSjg; spi – Pick-up station of si ; sdi – Drop-off station of si ; Ui – Set of pick-up stations in the upstream of si ; Di – Set of drop-off stations in the downstream of si ; pij – The probability that a moving request is picked up from spi and destined to sdj ; λi – Mean arrival rate of moving requests picked up from spi ; Λi – Mean arrival rate of moving requests dropped off to sdi . Parameters related to vehicles m – Quantity of vehicles in the system; ri – The probability that loaded vehicle drops off its load at sdi ; qi – The probability that an empty vehicle arriving at spi finds a waiting moving request; αdi – Rate of loaded vehicles’ arrival at sdi ; αpi – Rate of loaded vehicles’ arrival at spi ; εdi – Rate of empty vehicles’ arrival at sdi ; εpi – Rate of empty vehicles’ arrival at spi ; θi – Arrival rate of empty and loaded vehicles to stocker si ; pdi – The probability that the vehicle at spi− 1 is blocked by another vehicle at sdi ; ppi – The probability that the vehicle at sdi is blocked by another vehicle at spi ; βi – The probability that vehicle at sdi is empty when it is blocked by another vehicle at spi ; γi – The probability that vehicle at spi is empty when it is blocked by another vehicle at sdi+ 1 ; ti,i + 1 – The vehicle’s travel time from the pick-up station spi to the drop-off station sdi+ 1 ; ti – The vehicle’s travel time from the drop-off station sdi to the pick-up station spi , ti < ti,i + 1 ; tl – The vehicle’s picking-up and dropping-off time; tb – The mean time that vehicle is blocked by the vehicle in the downstream station;

4.3 Modified Markov Chain Model

83

τw – The probability that a loaded vehicle in the pick-up station spw gets through the shortcut connected with stocker si and enters the drop-off station sdw + k , τw = 1 − τw ; τw –The probability that a loaded vehicle in the pick-up station spw enters the drop-off station sdw + 1 . 3. Parameters related to the MMCM states R – State transition probability matrix of the MMCM; W – State set of the MMCM; v – State visited ratio vector of the MMCM, v = fvw g, w 2 W; vw – Visited ratio to state w of the MMCM; C – Mean cycle time between two visits to some reference state, e.g., ð1, p, eÞ; Tw – The time from the instant that the vehicle enters state w until the instant that it enters the next state. 4. System assumptions (1) Moving requests per period in each stocker is constant. (2) Moving request arrived according to the Poisson distribution process. (3) Each vehicle transports wafer lots based on the FCFS rule. (4) The loading time, unloading time and running speed of vehicles are deterministic values, and acceleration and deceleration of vehicles are ignored. (5) The stokers have enough storage capacities. (6) Vehicles travel independently in the AMHS, i.e., the correlation among the vehicles is not considered. (7) The probability that a vehicle is blocked by the downstream stocker is proportional to the number of vehicles in the AMHS. (8) Since each wafer lot dropped to a stocker is also picked up from the same stocker, the mean arrival rate of unloading requests and loading requests in each stocker is equal.

4.3.2 MMCM’s State Definition According to assumption 6, the process of m vehicles transporting wafer lots is equivalent to m independent processes of single-vehicle transporting wafer lots. According to assumption 2, the time of each single-vehicle picking-up wafer lot is stochastic. That is, the state of each single-vehicle transporting wafer lot also changes randomly. Therefore, the process of each single-vehicle transporting wafer lot is analyzed as a Markov chain process, which is modeled by using the MMCM. As only those pick-up and drop-off stations are considered, a state of the MMCM for the Interbay AMHS is defined by the three-tuple structure: MMCM = fS, Kp, Ksg = fðsi , i = 1, ..., nÞ, ðp, dÞ, ðe, f , s, bÞg

(4:1)

84

4 Analysis of automated material handling systems in SWFSs

where Kp is the pick-up and drop-off station set where vehicles locate. Ks is the state set of vehicles at pick-up and drop-off stations. p, d indicate vehicles’ location at pickup station and drop-off station, respectively. e, f , s, b mean that the state of a vehicle at pick-up and drop-off stations is empty running, loaded running, receiving pickingup/dropping-off service and blocked, respectively. S is denoted as the set of stockers, and si is the ith stocker in the Interbay system. According to the definition of the state, the whole set of states for the vehicle at the drop-off station includes ði, d, eÞ, ði, d, f Þ, ði, d, sÞ and ði, d, bÞ, which stand for empty running, loaded running, dropping-off service and blocked, respectively. The set of states for the vehicles at the pick-up stations is the same. Therefore, for the whole Interbay system with n stockers, the set of system states for the MMCM can be denoted as W = fð1, d, eÞ, ð1, d, f Þ, ð1, d, sÞ, ð1, d, bÞ, ..., ðn, p, eÞ, ðn, p, f Þ, ðn, p, sÞ, ðn, p, bÞg (4:2) The number of system states is 2 × 4 × n = 8n.

4.3.3 Modeling Process of the MMCM The modeling process of the AMHS consists of the following steps. First, the state transition probability of the MMCM for a single-vehicle based material handling process is analyzed and the transition matrix R is derived. Second, the MMCM’s steady state is analyzed and the visit ratio model of each state in the MMCM is inferred. Third, vehicle blocking probabilities in the AMHS are analyzed and the system blocking probability model is built. Fourth, the stability condition of the AMHS is analyzed and the system stability condition model is constructed. Fifth, the transition time of each state in the MMCM is analyzed and the mean cycle length C between two states is obtained. Finally, on the basis of the visit ratio model of each state in the MMCM, system blocking probability model, system stability condition model and the mean cycle length C, the stability state equations of MMCM are built and the visit ratio of each state in the MMCM can be solved. The performance analytical indexes can be expressed by the visit ratio of each state in the MMCM.

4.3.3.1 Transition Probabilities The transition probability refers to a change in probability between states in the MMCM. Consider an Interbay system with n stockers; the state transition diagram is partially described in Figure 4.2. The detailed process of state changing and state transition analyzing is illustrated as follows. Denote that there is an empty vehicle that approaches the drop-off station sdi , i.e., the state of a model is ði, d, eÞ. The transitions from this state depend on whether the vehicle is blocked by the vehicle at the pick-up station spi . If the vehicle is blocked by the vehicle at the pick-up station

85

4.3 Modified Markov Chain Model

ri+1 d 󰜏i pi+1

d d (󰜏i pi+1 + 󰜏i pi+k ) d

pi

...

(i, d, b) βi

(i, d, e) pid

... (i, p, s)

qi

d 󰜏i pi+k

pid+1 (i, p, e)

d qi pi +1

d

qi pi +1

(i+1, d, e) 𝛾i

pid+1

(i+1, d, s)

...

(i+1, d, b)

...

p

(i+1, d, f )

ri+1 pi+1

𝛾i󰜏i +1 (i, p, b)

1 p

ri+1 pi+1

(i+1, p, f )

...

𝛾i󰜏i +1 (i+k, d, f )

...

(i+1, d, b)

...

βi+1 (i+1, p, e)

...

(i, p, b)

Figure 4.2: The state transition diagram of the MMCM for Interbay system.

spi , then the system enters the state ði, d, bÞ, which happens with probability ppi .

Otherwise, if there is no vehicle at the pick-up station spi , then the system enters

the state ði, p, eÞ, which happens with probability 1 − ppi . When the vehicle enters the

pick-up station spi , the system state changes from ði, p, eÞ to ði, p, sÞ, ði, p, bÞ or ði + 1, d, eÞ according to corresponding probability. (1) If the vehicle finds a wafer lot waiting at the pick-up station spi , then the vehicle starts picking up wafer lot and the system enters the state ði, p, sÞ, which happens with probability qi . (2) If the vehicle does not find a load waiting at spi and the vehicle is blocked by a vehicle at the dropoff station sdi+ 1 , then system the enters the state ði, p, bÞ with probability qi pdi+ 1 . (3) If the vehicle does not find a load waiting at spi and there are no vehicles at the drop-off

station sdi+ 1 , then the system enters the state ði + 1, d, eÞ, which happens with probability qi pdi+ 1 . The transition probability of other states of the system can be analyzed in the same way. Based on these probabilities, the transition matrix R, which specifies the movement of the vehicle between the states of a system, can be identified, as described in Figure 4.3. For the Interbay system with n stockers, the dimension of the transition matrix R is 8n × 8n. In the transition matrix R, τi presents the probability of a loaded vehicle in the pick-up station spi which passes through the shortcut connected with stocker si and enters the drop-off station sdi+ k . Denote that the Interbay n systemo has l shortcuts, whose starting point correspondsnto the stocker oset sh1 , ..., shl and terminal point corresponds to the stocker set sh1 + k1 , ..., shs + kl . Then if i = h1 , ..., hl , 0 < τi < 1. Otherwise, if i ≠ h1 , ..., hl , τi = 0.

86

.. .

1df 1ds 1de 1db 1pf 1ps 1pe 1pb

r1

p r1 p1 r1 p1p p1p p1p

f k)d s

k)d

(i +

(i +

1)d s 1)d (i + e 1)d b (i +

(i +

ipb (i +

...

...

p1p

1

p1p

β1

β1 p2d p2d

q1

q1 p2d ...

...

...

...

...

...

...

...

...

ipf ips ipb ipe (i + 1)df (i + 1)ds (i + 1)de (i + 1)db .. .

ips ipe

...

ipf

1p f 1p s 1p e 1p b

1d b

1d e

1d

f 1d s

1)d f

4 Analysis of automated material handling systems in SWFSs

qi

.. .

(i + k)df

Figure 4.3: The state transition matrix R of the MMCM for Interbay system.

4.3.3.2 Steady-State Analysis Since the vehicle moves continuously on the closed-loop Interbay system, the state number of the MMCM is finite, e.g., the state number of the Interbay system with n stockers is 8n. Denote v = fvw g, w 2 W, where vw means the steady-state probability of the visited ratio to state w and W is the state set of the MMCM. Without loss of generality, set vð1, p, eÞ = 1. Let R be the state transition probability matrix of the MMCM. For a finite state, positive recurrent Markov chain, the steady-state probabilities can be uniquely obtained by solving the square system of equations. R  v=v

(4:3)

vð1, p, eÞ = 1

(4:4)

Let Rij (i, j = 1,...,8n) be the element of the transition matrix R, which denotes the transition probability from state i to state j in the transition matrix. Then Rij can be represented as follows:   y6      y ð 1 − y3 Þ  p y4 p ð 1 − y4 Þ y5 ð 1 − y5 Þ y2 ð1 − y2 Þ y1 ð1 − y1 Þ d 3 d Rij = qi . qi . ri . ri . pi . pi . τi . τi . pi . pi (4:5)

87

4.3 Modified Markov Chain Model

where y1 , y2 , y3 , y4 , y5 and y6 are the binary variables, y1 , y2 , y3 , y4 , y5 , y6 2 {0,1}. The Rij is related with vectors q = fqi gði = 1, ..., nÞ, r = fri gði = 1, ..., nÞ,





pd = pdi ði=1, ..., nÞ, τ= fτi gði=h1 , ..., hl Þ, β = βi ði=1, ..., nÞ, γ = γi ði=1, ..., nÞ and

pp = ppi ði=1, ..., nÞ. According to the definition of probabilities βi and γi , the expressions of βi and γi can be described as follows: βi = εpi =ðεpi + αpi Þ

(4:6)

  γi = εdi = εdi + αdi

(4:7)

From eqs (4.1) and (4.2), the visit ratio to each state can be inferred as follows: vði, d, eÞ = qi − 1 ppi vði − 1, p, eÞ + γi − 1 vði − 1, p, bÞ

(4:8)

8 > if ði − 1Þ = h1 , ..., hl τi − 1 ðpdi vði − 1, p, f Þ + γi − 1 vði − 1, p, bÞ + pdi vði − 1, p, sÞ Þ > > > > > > d d > if ði − 1Þ ≠ h1 , ..., hl , h1 + k , ..., hl + k > < pi vði − 1, p, f Þ + γi − 1 vði − 1, p, bÞ + pi vði − 1, p, sÞ vði, d, f Þ =

pdi ðvði − 1, p, f Þ + vði − 1, p, sÞ Þ + γi − 1 vði − 1, p, bÞ + τi + k ðpdi ðvði + k, p, f Þ + vði + k, p, sÞ Þ > > > > > > else + γi + k vði + k, p, bÞ Þ > > > :

( νði, p, bÞ =

vði, d, bÞ = ppi ðvði, d, sÞ + vði, d, eÞ Þ + ri ppi vði, d, f Þ

(4:9) (4:10)

vði, d, sÞ = ri vði, d, f Þ

(4:11)

vði, p, eÞ = ppi vði, d, eÞ + ppi vði, d, sÞ + βi vði, d, bÞ

(4:12)

 vði, d, kÞ vði, p, f Þ = ri ppi vði, d, f Þ + β i

(4:13)

  τi pdi+ 1 + τi pdi+ k νði, p, f Þ + νði, p, sÞ + qi pi + 1 d νði, p, eÞ 



pdi+ 1 νði, p, f Þ + νði, p, sÞ + qi pdi+ 1 νði, p, eÞ νði, p, sÞ = qi νði, p, eÞ

if i = h1 , ...h1 (4:14) if i = h1 , ...h1 (4:15)

In eq. (4.15), some of these probability variables are unknown: specifically, the picking-up/dropping-off probability variables q = fqi g and r = fri g, i = 1, ..., n,



the blocking probabilities variables pd = pdi and pp = ppi , i = 1, ..., n, and the shortcut-selecting probabilities variables τ = fτi g, i = h1 , ..., hl . The probability

88

4 Analysis of automated material handling systems in SWFSs

variables r, q, τ can be obtained by system’s stability condition analyzing, and the probability variables pd , pp can be obtained based on system blocking probability analysis.

4.3.3.3 System Blocking Probabilities System blocking probability refers to the probability that the vehicle in a pick-up/ drop-off station is blocked by the vehicle of a downstream station. In the Interbay system, the distance between the drop-off station and the pick-up station of a stocker is less than the distance between stockers. Hence, a vehicle at the drop-off station sdi cannot be blocked by a downstream vehicle at the pick-up station spi unless this downstream vehicle is traveling toward the pick-up station (i.e., the system state is in ði, p, eÞ or ði, p, f Þ). However, a vehicle at the pick-up station spi will be blocked when there is a vehicle at the pick-up station sdi+ 1 (i.e., the system state is in (i + 1, d, + )). As the blocking probabilities increase linearly with vehicle quantity according to system assumption 7 in Section 4.3.1, the blocking probability variables pdi and ppi can be estimated as follows: pdi = ðm − 1Þ

vði, dÞ jRj P j=1

ppi = ðm − 1Þ

, ∀sdi , i = 1, ..., n

ðvði, dÞ + vði, pÞ Þ

vði, p, sÞ + vði, p, bÞ + vði, p, kÞ jRj P j=1

(4:16)

, ∀spi , i = 1, ..., n

(4:17)

ðvði, dÞ + vði, pÞ Þ

As the Interbay system runs stably, the following stability conditions should be satisfied. (1) For each stocker si , i = 1, ..., n, the mean arrival rate that moving requests dropped off to station sdi should be equal to the drop-off ratio of loaded vehicles at the current station. (2) For each stocker si , i = 1, ..., n, the mean arrival rate that moving requests picked up from station spi should be equal to the pickup ratio of empty vehicles at the current station. (3) According to from-to-table (it is a conveying relation table where the transition ratios between stockers in the Interbay system are described) of the Interbay system, vehicles in the Interbay system should pass through each shortcut by certain probabilities. Based on system stability conditions, the probability variables r, q and τ can be obtained. ① The probability variables r According to system stability conditions, for each stocker si ,i = 1, ..., n, the mean arrival rate of moving requests dropped off at station sdi , Λi , which is equal to the drop-off ratio of loaded vehicles at the current station, which is the product

89

4.3 Modified Markov Chain Model

s1 s1d

sw s1p

...

sw+1

sud

sup

d su+1

p su+1

Material handling zone 1 (named Z1)

snp

snd

...

shd

shp

Material handling zone 2 (named Z2)

p ... su+k

sn

sh

p

d su+k

d

su+k–1

sw+k

p ... sh+1

su+k–1

sw+k–1

d sh+1

sh+1

Figure 4.4: The Interbay material handling system with single-shortcut.

of the rate of loaded vehicles’ arrival at sdi , αdi and the probability that loaded vehicle drops off its load at sdi , ri . Then the vehicle’s picking up probability ri can be described as Λi = ri  adi ) ri = Λi =adi

(4:18)

adi can be obtained based on material handling flow analyzing of the Interbay system. For example, there is an Interbay system with single shortcut (Figure 4.4.), which can be separated as two material handling zones based on shortcut, labeled Z1 (i.e., the stocker set is fs1 , .., su , su + k , ..., sn g) and Z2 (i.e., the stocker set is fsu + 1 , .., su + k − 1 g). Assume that a loaded vehicle runs in Z1 and arrives at station spu ; then the vehicle will decide whether it should enter in Z2 according to the destination stocker of the wafer lot on the vehicle. If the destination stocker belongs to Z2 , then the loaded vehicle selects to enter in Z2 . Otherwise, if the destination stocker belongs to Z1 , the loaded vehicle runs in the current zone continuously. The equation of the parameter adi can be described as

adi =

8 j n P iP −1 P > > > λ p + > j jr > > j=1 > > j=i+1 r=i > < > > > > > > > > > :

n P

u +P k−1

j = i, j ≠ ½u + 1, u + k − 1

r=i

n P r=1 r ≠ ½j, i − 1

λj pjr +

λj pjr , ∀si 2 fs1 , .., su , su + k , ...., sn g iP −1

n P

j=u+1

r = 1, r ≠ ½j, i − 1

λj pjr +

j u +P k−1 P j=i+1 r=i

λj pjr , ∀si 2 fsu + 1 , .., su + k − 1 g

(4:19) The expression of adi in the Interbay system with multi-shortcut can be inferred by the same way. Since the SWFS has the characteristic of multi re-entrant, the wafer lot dropped off at station sdi will be picked up at station spi later. So, the mean arrival

90

4 Analysis of automated material handling systems in SWFSs

rate of moving requests picked up from spi ; λi equals the mean arrival rate of moving requests dropped off at sdi , Λi , i.e., Λi = λi . ② The probability variables q q = fqi g, i = 1, ..., n, denotes the probability that a loaded vehicle picks up a waiting moving request at sdi . According to system stability conditions, the mean arrival rate

of moving requests picked up from the station spi ; λi equals the pick-up ratio of an empty vehicle at the current station, which is the product of the rate of empty vehicles arriving at spi , εpi and qi . Then qi can be described as follows: λi = qi  εpi ) qi =

λi , εpi

∀i = 1, ..., n

(4:20)

In the mean cycle time C, the visit ratio that any vehicle enters the system state ði, p, eÞ εp C is vði, p, eÞ , which can be described as follows: vði, p, eÞ = i , where m is the vehicle m number in the Interbay system. Then, qi can be inferred as qi =

λi C vði, p, eÞm

(4:21)

The value of the mean cycle time C will be solved in Section 4.3.3.4. ③ The probability variables τ τ = fτi g,i = h1 , ..., hl , denotes the probability that a loaded vehicle at the pick-up station spi ði = h1 , ..., hl Þ enters the downstream drop-off station sdi+ 1 . For an Interbay system with a single shortcut (Figure 4.4.), assume that there is a moving request at the pick-up station spu in Z1 . If an empty vehicle arrives at spu and picks up the load, then the loaded vehicle will choose the running path according to the destination stocker of wafer lots on the vehicle. If the destination stocker belongs to Z1 , then the loaded vehicle will pass through shortcut and run along paths fsu ! su + k !    ! sn !   g. Otherwise, if the destination stocker belongs to Z2 , the vehicle will enter Z2 and run along paths su ! su + 1 !    ! sh !   g. The equation of τu can be described as follows: u n iP −1 P P ð pi j + pi j Þ +

τu =

i=1 j=u+k u P i=



n P

j=u+1

j=1

pi j +

iP −1 j=1

n P

i P

i=u+k+1 j=u+k

pi j Þ +

n P

i P

i=u+2 j=u+1

pi j (4:22)

pi j

The expression of τu in an Interbay system with multiple shortcuts can be inferred by the same way.

4.3 Modified Markov Chain Model

91

4.3.3.4 The Mean Cycle Time C C is the mean cycle time between two successive visits to some reference state, e.g., ð1, p, eÞ. Denote Tw as the time from the instant the vehicle enters state w until the instant it enters the next state. According to the definition of the mean cycle time, the expression of C can be described as C= =

jWj P

vw EðTw Þ

w=1 n  P i=1

vði,d,eÞ EðTði,d,eÞ Þ + vði,d,f Þ EðTði, d, f Þ Þ + vði,d,bÞ EðTði,d,bÞ Þ + vði,d,kÞ EðTði,d,kÞ Þ

(4:23)

+ vði,d,sÞ EðTði,d,sÞ Þ + vði,p,eÞ EðTði,p,eÞ Þ + vði, p, f Þ EðTði,p,f Þ Þ + vði,p,bÞ EðTði,p,bÞ Þ

+ vði,p,kÞ EðTði,p,kÞ Þ + vði,p,sÞ EðTði,p,sÞ ÞÞ where EðTw Þ is the expected mean state time, which can be determined based on the state transition probabilities. The expected mean state time of each state is described as follows. (1) Consider the state ði, d, sÞ=ði, p, sÞ, which means vehicle dropping off/picking up wafer lot at the station sdi =spi ; the time that the vehicle spends in these states are the vehicle’s picking-up and dropping-off times tl . (2) For the state ði, d, f Þ=ði, d, eÞ, which means a loaded/empty vehicle arriving at some drop-off station sdi , the time that the vehicle spends in these states is the traveling time of a loaded/empty vehicle from the pick-up station spi− 1 to the drop-off station sdi , ti − 1, i . (3) The time that the vehicle spends in the state ði, p, f Þ=ði, p, eÞ is the traveling time of a loaded/empty vehicle from the drop-off station sdi to the pick-up station spi , ti . (4) The state ði, d, bÞ means a vehicle blocked at the drop-off station sdi , which happens when there is another vehicle picking up wafer lots at the station spi . The time that a vehicle spends in this state is related to its previous states. (5) Consider the state ði, p, bÞ, which means a vehicle is blocked at the pick-up station spi , and the time that the vehicle spends in this state is influenced by the states in the downstream station sdi+ 1 . The expressions of the expected transition time for each state of a vehicle at sdi and spi are derived as follows:

EðTði, d, bÞ Þ =

EðTði, d, f Þ Þ = ti − 1, i

(4:24)

EðTði, d, eÞ Þ = ti − 1, i

(4:25)

βi × tl × vði, d, eÞ β × tl + i 2 × ðvði, d, eÞ + vði, d, sÞ Þ 2

(4:26)

EðTði, d, sÞ Þ = tl

(4:27)

92

4 Analysis of automated material handling systems in SWFSs

EðTði,p,f Þ Þ = ti

(4:28)

EðTði,p,eÞ Þ = ti

(4:29)

  E Ti,p,b =     νði + 1,d,f Þ + νði + 1,d,eÞ ðti,t + 1 − ti Þ + νði + 1,d,sÞ ðti,t + 1 − ti Þ + tl   νði,p,f Þ + ðti,t + 1 − ti Þνði,p,sÞ νði + 1,d,f Þ + νði + 1,d,eÞ + νði + 1,d,sÞ γi ðνði,p,f Þ + νði,p,sÞ Þ × 2     νði + 1,d,f Þ + νði + 1,d,eÞ ðti,t + 1 − ti Þ + νði + 1,d,sÞ ðti,t + 1 − ti Þ + tl   + γi νði + 1,d,f Þ + νði + 1,d,eÞ + νði + 1,d,sÞ × 2 (4:30)

EðTði, p, sÞ Þ = tl

(4:31)

According to analytical results from Sections–4.3.3.4, the stability state equations of the MMCM are obtained and described as follows:

pdi =

(4:32)

vðs1 , p, eÞ = 1

(4:33)

ðm − 1Þvðsi , d, *Þ jW j P j=1

ppi =

R  v=v

, ∀sdi , i = 1, ..., n

ðm − 1Þvðsi , p, sÞ + vðsi , p, bÞ j Wj

P

j=1

(4:34)

ðvðsi , d, *Þ + vðsi , p, *Þ Þ

, ∀spi , i = 1, ..., n

(4:35)

ðvðsi , d, *Þ + vðsi , p, *Þ Þ

qi =

Ct =

λi Ct vðsi , p, eÞ m

jW j X w=1

vw EðTw Þ

(4:36)

(4:37)

4.3 Modified Markov Chain Model

93

The quantity of stability state equations described above is jW j + 3n + 1. The number of variables in these equations is also jW j + 3n + 1, which includes jW j steady-state visit ratio variables vw ðw 2 WÞ, n picking-up probability variables qi ði = 1, ..., nÞ, n vehicle-blocking probability variables ppi ði = 1, ..., nÞ and pdi ði = 1, ..., nÞ, and a mean cycle time variable Ct. Proposition 1. For the system with non-linear equations given in above equations, if the Interbay AMHS can handle all moving requests within the planning horizon, i.e., the Interbay AMHS is stable, then there exist unique values for the variables qi ði = 1, ..., nÞ, ppi ði = 1, ..., nÞ, pdi ði = 1, ..., nÞ and Ct, and these values provide the sole solution to the steady-state visit ratio vector v and the state transition probability matrix R in the MMCM. 4.3.4 Model Validation 1. Experimental data This section presents a numerical experiment study to demonstrate the effectiveness of the proposed MMCM analytical approach. The system data used for numerical experiments in this study were shared by a semiconductor manufacturer in Shanghai, China. To protect the confidentiality of the manufacturer’s information, some system data presented in this chapter have been modified and are for demonstrative purpose only. There are 14 stockers in the Interbay material handling system, and each stocker connects with an Intrabay system (Figure 4.5). The distance between adjacent stockers is 20 m, the distance between the pick-up port and the drop-off port of a stocker is 5 m and the Interbay rail loop is 320 m long with two shortcuts. Each automated vehicle can load only one wafer lot at a time, and the vehicle’s picking-up and dropping-off time at a stocker is 10 s. Two types of wafer lot (jobs A and B) are being handled in the system and the release ratio of jobs A and B is 1:1. The system’s impact factors considered in the

S1

S2

S3

S4

S5

S6

S7

S14

S13

S12

S11

S10

S9

S8

Drop-off station Pick-up station

Vehicle Stocker

Figure 4.5: The layout of the Interbay material handling system.

94

4 Analysis of automated material handling systems in SWFSs

Table 4.1: Investigated factors. Factor

Levels

Vehicle quantity System’s release ratio Vehicle’s speed

3–15 vehicles 9,000 wafer/month, 13,500 wafer/month, 15,000 wafer/month 1 m/s, 1.5 m/s

experimental study include vehicle quantity, system’s release ratio and vehicle’s speed, as listed in Table 4.1. Totally, 72 scenarios are designed for different combinations of the selected factors. With the actual data of the Interbay material handling system, the basic data and key system parameters of the MMCM in each scenario are identified, including from-to-table of the Interbay system, mean arrival rate of moving requests picked up from spi , λi , mean arrival rate of moving requests dropped off to sdi , Λi , rate of loaded vehicles’ arrivals at sdi , αdi , the dropping-off probability ri and the shortcutselecting probability τi . For instance, in the scenario where system’s release ratio = 13,500 wafer/month and vehicle’s speed = 1 m/s, the from-to-table and the key system parameters of the MMCM are calculated and shown in Tables 4.2 and 4.3, respectively. 2. Comparison and discussion Based on the basic data and the values of key system parameters, the MMCM analytic model is comprehensively compared with discrete event simulation model in each scenario. A total of three performance indexes, system’s throughput capability, vehicle’s mean utilization ratio and mean arrival time interval of empty vehicle, are compared. The eM-Plant simulation software is adopted to construct Interbay’s simulation model, in which moving requests arrive at the stocker according to Poisson process. The simulation model in each scenario is executed with three replications of 60 days each. The performance analysis and comparison results of the MMCM analytic model and the simulation analytic model are partially demonstrated in Tables 4.4–4.6. Table 4.4 shows that, in those scenarios that the arrive ratio of moving requests = 56 lot/h and vehicle speeds = 1 m/s, the mean relative error between the MEMCM and the simulation model in terms of empty vehicle’s mean arrival time interval, throughput capacity and vehicle’s mean utilization ratio is 3.75%, 2.88% and 1.86%, respectively. Table 4.5 illustrates that, in those scenarios that arrive ratio of moving requests = 84.5 lot/h and vehicle speeds = 1 m/s, the mean relative error between the MMCM and the simulation model in terms of empty vehicle’s mean arrival time interval, throughput capacity and vehicle’s mean utilization ratio is 3.6%, 2.87% and 10.0%, respectively. Table 4.6 shows that, in the scenario that the number of vehicles = 7, vehicle speeds = 1 m/s and the arrive ratio of moving requests = 56 lot/h, and the

S1

0 0 0.5 0.5 1 0.5 1 0 2 0.5 0 0 1 0

Stocker ID

S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 S12 S13 S14

1.5 0 0 0.5 0 0 0 1 0 0.5 1.5 1 0 1.5

S2 0.5 0.5 0 0 0 0.75 0.5 0 0 0 1.5 0.5 1.5 2

S3 1 0.5 0 0 0 0 0 0.5 0 1 1 0.5 0 0.5

S4 0 0.5 0 0 0 0.5 0 0 1 0.5 0 0.5 0 0.5

S5 1 1 1 0 0 0 0.5 0.25 0 0.5 0.5 0 0 0

S6 0 0 1 0.5 0 0.5 0 0 0.5 0.5 1 0.5 0 0

S7 0.5 0.5 1 0 0 0 0 0 0 0 0 1.5 0 1

S8 1 0.5 0.25 1 0 0 1 0 0 0 0.75 1 1 1

S9 0 0.5 1 0 0 0.5 0 0.75 0 0 1 0.5 2.5 0

S10

Table 4.2: The from-to-table of the Interbay system (system’s release ratio = 13,500 wafer/month, vehicle speed = 1 m/s).

0.5 2 1 0 0 1 0 0.5 1 1.25 0 0 0 0

S11

0 0 0.5 0.5 1 1 0.5 0.5 1 1 0 0 0 0

S12

0.5 1 0.5 1 1 0 0.5 0.5 1 0 0 0 0 0

S13

0.5 0.5 1 1 0.5 0 0.5 0.5 1 1 0 0 0 0

S14

4.3 Modified Markov Chain Model

95

0.349 /

0.304 /

adi ri τw

7.5 7.5 21.5

7 7 23

λi Λi

S2

S1

Variable

0.190 /

7.75 7.75 40.75

S3

0.123 /

5 5 40.75

S4

0.086 0.515

3.5 3.5 40.75

S5

0.241 /

4.75 4.75 19.75

S6

0.228 /

4.5 4.5 19.75

S7

0.228 /

4.5 4.5 19.75

S8

0.380 /

7.5 7.5 19.75

S9

0.166 /

6.75 6.75 40.75

S10

0.178 /

7.25 7.25 40.75

S11

Table 4.3: The key system parameter values of the MMCM (system’s release ratio = 13,500 wafer/month, vehicle speed = 1 m/s).

0.147 0.460

6 6 40.75

S12

0.261 /

6 6 23

S13

0.283 /

6.5 6.5 23

S14

96 4 Analysis of automated material handling systems in SWFSs

3 4 5 6 7 8 9 10 11 12 13 14 Mean value

Vehicle number

200.7 119.5 85.6 67.0 55.2 47.2 41.3 36.8 33.2 30.4 28.0 26.1 /

MEMCM

212.3 122.2 86.0 66.5 54.2 45.9 39.8 35.2 31.5 28.6 26.2 24.2 /

Simulation model 5.8 2.3 0.5 0.7 1.8 2.8 3.6 4.3 5.1 5.9 6.4 7.3 3.75

Relative error (%)

Empty vehicle’s mean arrival time interval (s)

68,344 91,151 113,963 136,782 159,610 182,450 205,301 228,167 251,048 273,947 296,863 319,800 /

MEMCM 70,458 93,953 117,437 140,920 164,401 187,907 211,382 234,846 258,343 281,873 305,318 328,835 /

Simulation model 3.00 2.98 2.96 2.94 2.91 2.90 2.88 2.84 2.82 2.81 2.77 2.75 2.88

Relative error (%)

Throughput capacity (lot)

60.4 45.7 36.8 31.0 26.8 23.6 21.2 19.2 17.6 16.3 15.2 14.2 /

MEMCM

62.2 46.3 37.1 31.1 26.7 23.5 20.8 18.9 17.3 15.9 14.8 13.8 /

Simulation model

2.9 1.3 0.8 0.3 0.4 0.4 1.9 1.6 1.7 2.5 2.7 2.9 1.86

Relative error (%)

Vehicle’s mean utilization ratio (%)

Table 4.4: Performances comparison of the MMCM and the simulation-based model (arrive ratio of move requests = 56 lot/h, vehicle speeds = 1 m/s).

4.3 Modified Markov Chain Model

97

4 5 6 7 8 9 10 11 12 13 14 15 Mean value

Vehicle number

178.8 112.1 82.1 65.1 54.2 46.5 40.9 36.5 33.1 30.3 28.0 26.1 /

MEMCM 203.6 120.0 85.4 66.4 54.5 46.3 40.3 35.7 32.1 29.3 27.0 25.1 /

Simulation model 12.2 6.6 3.9 2.0 0.6 0.4 1.5 2.2 3.1 3.4 3.7 4.0 3.6

Relative error (%)

Empty vehicle’s mean arrival time interval (s)

91,245 114,092 136,948 159,818 182,704 205,607 228,531 251,477 274,447 297,444 320,470 343,525 /

MEMCM 94,013 117,509 141,010 164,507 188,015 211,514 235,015 258,524 282,001 305,526 329,035 352,530 /

Simulation model 2.94 2.91 2.88 2.85 2.82 2.79 2.76 2.73 2.68 2.65 2.60 2.55 2.87

Relative error (%)

Throughput capacity (lot)

68.5 55.3 46.4 40.1 35.4 31.7 28.8 26.4 24.4 22.7 21.3 20.1 /

MEMCM

63.5 50.9 42.6 36.7 32.2 28.8 26.1 23.8 22.0 20.4 19.2 18.1 /

Simulation model

7.9 8.6 8.9 9.3 9.9 10.1 10.3 10.9 10.9 11.3 10.9 11.0 10.0

Relative error (%)

Vehicle’s mean utilization ratio(%)

Table 4.5: Performances comparison of the MMCM and the simulation-based model (arrive ratio of move requests = 84.5 lot/h, vehicle speeds = 1 m/s).

98 4 Analysis of automated material handling systems in SWFSs

4.4 Analysis of AMHS Based on the MMCM

99

Table 4.6: Empty vehicle’s mean arrival time interval comparison between the MMCM and the simulation-based model. Stocker ID

S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 S12 S13 S14

Vehicles = 7, vehicle speeds = 1 m/s, arrive ratio of move requests = 56 lot/h

Vehicles = 7,vehicle speeds = 1 m/s, arrive ratio of move requests = 84.5 lot/h

MEMCM (s)

Simulation model (s)

Relative error (%)

MEMCM (s)

Simulation model (s)

Relative error (%)

55.5 55.2 54.8 56.2 57.0 56.0 55.9 55.7 53.8 54.2 54.0 54.7 55.2 55.4

53.7 53.4 53.5 54.7 55.6 54.9 55.2 55.1 53.4 53.9 53.4 54.4 54.3 54.0

3.24 3.26 2.37 2.67 2.46 1.96 1.25 1.08 0.74 0.55 1.11 0.55 1.63 2.53

65.5 64.9 64.1 67.2 68.9 66.6 66.4 66.0 62.1 63.0 62.5 64.0 65.1 65.5

65.2 64.6 64.6 67.7 69.6 68.0 68.6 68.3 64.6 65.8 64.6 66.7 66.4 65.8

0.46 0.46 0.78 0.74 1.02 2.10 3.31 3.48 4.03 4.44 3.36 4.22 2.00 0.46

scenario that the number of vehicles = 7, vehicle speeds = 1 m/s and the arrive ratio of moving requests = 84.5 lot/h, the mean relative error of empty vehicle’s mean arrival time interval between the MMCM and the simulation model is less than 4.44%. The results demonstrate that the MMCM analytical approach is feasible. To further evaluate the effectiveness of the MMCM analytical approach, the frequency of relative error percentages of the MMCM and the simulation model in all scenarios are analyzed. Figure 4.6. illustrates that 96% of all relative error values belong to [–8%, 10%]. Meanwhile, runtime comparison between the MMCM and the eM-Plant model-based analyzing methods is given in Table 4.7 and the results show that the computation efficiency of the MMCM is improved about 600 times than that of the eM-Plant model. It demonstrated that the proposed MMCM analytical approach performed reasonably well with acceptable error percentages and was an effective performance analyzing approach for the AMHS with shortcut and blocking.

4.4 Analysis of AMHS Based on the MMCM The stability state equations of the MMCM are solved quickly using Gauss iterative method [3, 4]. Based on the MMCM and its state visit ratio, the performance of the Interbay AMHS is analyzed and evaluated.

100

4 Analysis of automated material handling systems in SWFSs

Table 4.7: Runtime comparison between the MMCM and eM-Plant model based analyzing methods. Analyzing methods

Runtime (min)

eM-Plant model MMCM

≈60 ≈0.1

0.2 Relative error percentage

0.18 0.16 0.14 0.12 0.1 0.08 0.06 0.04 0.02 0 –15

–10

–5

0 5 10 15 20 The value of relative error (%)

25

30

35

Figure 4.6: The frequency of relative error percentages of the MEMCM and the simulation model.

4.4.1 The Overall Utilization Ratio and the Mean Utilization Ratio of a Vehicle The overall utilization ratio ρ refers to the ratio of times for a vehicle to stay in the handling wafer state, the loading state and the blocking state. P ρu =

w2X ∪ F ∪ K jW j P w=1

P

vw Tw =

w2X ∪ F ∪ K

vw Tw

C

vw Tw

The mean utilization ratio ρu means the ratio of times that a vehicle stays in the handling wafer state and the loading state. P ρu =

r2X ∪ F jW j P w=1

P

vw Tw

vw Tw

=

r2X ∪ F

vw Tw

C

(4:38)

where X is the set of states when the vehicle is picking up or dropping off. F is the set of states when the vehicle is loaded. K is the set of states when the vehicle is blocked.

4.4 Analysis of AMHS Based on the MMCM

101

4.4.2 The Mean Arrival Time Interval of an Empty Vehicle The mean arrival time interval of an empty vehicle Te refers to the mean cycle time for an empty vehicle to successively visit the pick-up station spi of the stocker si. The smaller the value of Te , the shorter the mean waiting time of the wafer lot at a stocker. Te can be calculated as follows: Te =

n n Tði, p, eÞ 1X 1 1X p = n i = 1 εi n i = 1 vði, p, eÞ

(4:39)

4.4.3 Expected Throughput Capability and Real Throughput Capability of an AMHS The expected throughput capability of an AMHS refers to the expected maximum number of wafer lots that can be handled in a given period. Denote Mov as the expected throughput capability of an AMHS; the expression of Mov is defined as n P

Mov = m  i = 1

ðvði, p, eÞ + vði, p, f Þ Þ nC

where vði, p, eÞ + vði, p, f Þ is the probability vehicle arriving at states ði, p, f Þ and ði, p, eÞ. The real throughput capability of an AMHS refers to the number of wafer lots that the AMHS can handle in a given period. Denote TH as the real throughput capability of an AMHS; the expression of TH is defined as n P

Mov = m 

i=1

ðvði, p, eÞ + vði, p, f Þ Þ nC

(4:40)

where n is the stocker quantity in an Interbay system and m is the number of vehicles in an Interbay system.

4.4.4 The Waiting Time of a Wafer The waiting time of a wafer Ri refers to the total time that a wafer stays in the stocker before it is loaded to the vehicle. Assuming that moving requests arrive at the stocker according to Poisson process, Ri is the expected waiting time of wafer x at the loading point spi and there are L lots at the loading point spi before wafer x arrives. Then the expected waiting time of wafer x is defined as

102

4 Analysis of automated material handling systems in SWFSs

Ri =

∞ X

Ri ðLÞPi ðLÞ

(4:41)

L=0

where Pi ðLÞ is the probability that there are L lots to be transported at loading point spi , L 2 Z + ; Ri ðLÞ is the expected waiting time of wafer x at the loading point spi when

there are L lots at the loading point spi . Ri ðLÞ consists of two parts: the expected

waiting time of the first lot at the loading point spi and the total waiting time of the remaining lots. Then Ri ðLÞ can be given by Ri ðLÞ = Ri ð0Þ + L  Tei

(4:42)

where Tei is the average time of the empty vehicle to the stocker. According to Little’s law: WIPsd = λi  Ri = i

∞ X

L  Pi ðLÞ

(4:43)

L=0

where λi is the arrival rate of moving requests at the loading point and WIPsd is i the expected WIP at the loading point spi . Then it can be further defined as Ri =

∞ P

ðRi ð0Þ + L  Tei Þ  Pi ðLÞ =

L=0

+ Tei

∞ P L=0

L

∞ P L=0

Pi ðLÞ = Ri ð0Þ + Tei

Ri ð0Þ  Pi ðLÞ + Tei

 λi  Ri ,

)

∞ P L=0

L  Pi ðLÞ = Ri ð0Þ

Ri = Ri ð0Þ=ð1 − Tei

(4:44)  λi Þ

According to the system’s hypothesis, the arrival time of an empty vehicle to the loading point spi is random, and the average time of the empty vehicle to the stocker is Tei . The computation formula of remaining waiting time, and the expected waiting time of the first lot can be denoted by EðRi ð0ÞÞ =

EðTei Þ + ðCT2 i + 1Þ 2

=

Tei 1 = p 2 2εi

(4:45)

where CT2 i is the variance of the cycle time vehicle arriving at the loading point spi . Then the expected waiting time of wafer x can be defined as

Ri =

Tei 2ð1 − Tei  λi Þ

(4:46)

4.4 Analysis of AMHS Based on the MMCM

103

4.4.5 Vehicle Blockage-Related Indicators Vehicle blockage-related indicators include vehicle blockage rate β, vehicle efficient utilization rate ρR , system’s material handling capability reduction rate ρmov and empty vehicle arrival time interval elongation rate ρT . Each performance indicator is presented as follows. 1. Vehicle blockage rate β. Vehicle blockage rate refers to the ratio of the vehicle blocking time to the overall handling time in the material handling process. The vehicle blockage rate β based on the MMCM can be calculated by P β=

r2B ∪ K j Rj P r=1

P

vr Tr =

vr Tr

r2B ∪ K

vr Tr

C

(4:47)

where K is the set of states in which a vehicle is fully loaded and B is the set of states in which vehicle e is unloaded and blocked. 2. Vehicle efficient utilization rate ρR . Vehicle efficient utilization rate mainly describes the influence of the vehicle blockage on the vehicle utilization rate, and ρR is the ratio of the effective utilization to the overall vehicle utilization, which can be given by ρR =

ρu × 100% ρ

(4:48)

3. System’s material handling capability reduction rate ρmov . System’s material handling capability reduction rate mainly describes the influence of the vehicle blockage on the whole handling capacity of an AMHS, and can be defined by

ρmov =

TH=ρu − Mov × 100% TH=ρu

(4:49)

4. Empty vehicle arrival time interval elongation rate ρT . Empty vehicle arrival time interval elongation rate mainly describes the influence of the vehicle blockage on the average arrival time interval of an empty vehicle, and can be calculated by

104

4 Analysis of automated material handling systems in SWFSs

 ρT =

n×β×C =Te × 100% n P m × vði, p, eÞ

(4:50)

i=1

4.4.6 Case Study In this section, practical data collected from a 300 mm wafer fabrication factory in Shanghai are analyzed. The AMHS in the factory consists of one Interbay subsystems and 22 Intrabay subsystems. They are connected by stockers, with a total of 496 devices distributed in 22 Intrabay subsystems. The analysis process is presented as follows: 1. Input system handling track (including shortcut track) layout information, stocker location and loading/unloading port position information, as well as vehicle information. 2. Set parameters of the modified Markov model. This includes set wafer arrival rate, processing route of different wafers and their corresponding processing areas in the modified Markov model. 3. Construct modified Markov model. According to the input parameters of the modified Markov model, establish wafer arrival rate model of each stocker, system conservation and stability condition model, state transition probability model and vehicle blockage probability model, and obtain the steady-state probability of the modified Markov model. 4. Analyze the running characteristics of an AMHS based on the modified Markov model including the impacts of vehicle number, vehicle speed and vehicle blockage on the system’s performance. The results are shown in Figures 4.7–4.9.

Vehicle utilization rate (100%)

90 80 70 60 50 40 30 20 10 0 4

5

6

7

8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 Vehicle number

Figure 4.7: Impact of the vehicle number on the vehicle utilization rate.

4.5 Conclusion

105

2.4 Vehicle overall handling capability × 105 (lot)

2.2 2.0 1.8 1.6 1.4 1.2 1.0 0.8 0.6 4

5

6

7

8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 Vehicle number

Empty vehicle arrival time interval (s)

Figure 4.8: Impact of the vehicle number on the vehicle overall handling capability.

240 220 200 180 160 140 120 100 80 60 40 20 0 4

5

6

7

8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 Vehicle number

Figure 4.9: Impact of the vehicle number on the empty vehicle arrival time interval.

4.5 Conclusion AMHS is indispensable and critical in 300 mm SWFSs. In order to analyze and evaluate the performances of AMHS with shortcut and blockage effectively and efficiently in system’s design stage, an MMCM has been proposed. With real production data, the proposed MMCM has been compared with the simulation model. The results have demonstrated that the proposed MMCM and the simulation approach have small relative errors in terms of system’s throughput capability, vehicle’s mean utilization ratio and empty-vehicle’s mean arrival time interval. The 96% of all relative error percentages range from –8% to 10%. In the aspect of computational

106

4 Analysis of automated material handling systems in SWFSs

efficiency, the execution time of the simulation model is about several hundred times more than that of the MMCM. It means that the proposed MMCM is an effective and efficient modeling approach for analyzing the large-scale and complex AMHS’s performances with shortcut and blockage.

References [1] [2] [3] [4]

Sematech Report. Global joint guidelines for 300 mm semiconductor factories I300I/J300. Available at: International SEMATECH website. 1997, http://www.sematech.org. Wu, L.H.Research on intelligent scheduling technologies of AMHS in semiconductor wafer fabrication system, Shanghai Jiao Tong University, 2011. Sampoorna, M., Bueno, J.T. Gauss-Seidel and successive overrelaxation methods for radiative transfer with partial frequency redistribution. Astrophysical Journal, 2010, 712(2): 1331–1344. Chatterjee, C., Roychowdhury, V.P., Chong, E.K.P. A nonlinear Gauss-Seidel algorithm for noncoplanar and coplanar camera calibration with convergence analysis. Computer Vision and Image Understanding, 1997, 67(1): 58–80.

5 Scheduling methods of automated material handling systems in SWFSs In order to respond quickly to changes in market demands, improve on-time delivery rate and shorten the production cycle time, it is necessary to schedule wafer fabrication systems and AMHS. The overall optimization objectives are achieved by reducing transport time, waiting time and WIP quantity. Scheduling in AMHSs refers to meeting the material handling constraints and arranging the sequence and time of wafer carriers handled by vehicles in order to optimize the material handling performance. The scheduling problem in AMHSs is a typical NPhard problem due to its large-scale [1], stochastic, real-time and multi- objective features. It is ineffective to solve such a complex problem by using traditional mathematical programming methods. In addition, the dynamic scheduling problem of complex production processes in the large-scale uncertain environment cannot be solved by using the traditional scheduling theory. Therefore, the scheduling problem in AMHSs is intensively studied by domestic and foreign scholars, using the common methods including heuristic rules, operation research methods and artificial intelligence algorithms.

5.1 Heuristic Rules-Based Scheduling Methods The heuristic approach refers to a method of finding an optimal schedule based on empirical rules. It solves problems by using past experiences and proven methods rather than using determinate steps to find the answer. The heuristic method reduces the number of attempts in the limited search space and solves the problem quickly [2]. Many major discoveries of scientists are obtained by using simple heuristic rules. In the information processing theory of the cognitive psychology, heuristic is an important way to solve the problem of human thinking. In artificial intelligence, heuristic methods are often used to design computer programs in order to simulate human thinking activities. It has been proved that this is an effective way to solve practical problems.

5.1.1 Heuristic Rules The method based on the single heuristic rule is to use a kind of heuristic rule as the optimization strategy. It is usually applied to optimize the single performance index of the AMHS, such as the shortest wafer processing cycle, the shortest wafer waiting time, the minimum number of wafer products and the highest machine utilization https://doi.org/10.1515/9783110487473-005

108

5 Scheduling methods of automated material handling systems in SWFSs

Table 5.1: Dispatching rules. RULE

DATA

ENVIRONMENT

1 2

SIRO ERD

_ rj

3 4 5

EDD MS SPT

dj dj pj

6 7 8 9 10 11 12 13 14 15

WSPT LPT SPT-LPT CP LNS SST LFJ LAPT SQ SQNO

wj, pj pj pj pj, prec pj, prec sj,k Mj pij _ _

_   P   1  rj  Var Cj − rj =n 1 k Lmax 1 k Lmax P P Pm k Cj ; Fm j pij = pj  Cj P ωj Cj Pm k Pm k Cmax Fm j block, pij = pj j Cmax pm j prec j Cmax pm j prec j Cmax 1 j sjk j Cmax pm j Mj j Cmax O2 k Cmax P Cj Pm k Jm k γ

rate. However, it is difficult to use the single heuristic rule-based method to optimize the multi-objective AMHS. Commonly used heuristic rules are presented in Table 5.1. For a number of heuristic rules, each rule has a different nature; hence, there is no uniform measure. There is a contradiction between the rules; the objective value under one rule is improved while the objective value under another rule is compromised. The process that uses a variety of prioritized strategies and considers various heuristic rules to optimize is regarded as a heuristic rule composition. The method based on the compound heuristic rules that consider multiple system parameters, such as idle vehicle delivery time, wafer waiting time, the number of wafer cards waiting to be delivered, the distance between the wafer lot and the vehicle, and the current vehicle location, is able to optimize the multi-objective AMHS.

5.1.2 Heuristic Rules Applied in AMHS Scheduling For the scheduling problem in AMHSs, the system parameters such as the idle vehicle delivery time, the wafer waiting time and the number of the wafer lots waiting for transportation are taken into consideration. According to the actual demand of the manufacturing system, the heuristic search is implemented to quickly obtain the search results in order to determine the material handling program in AMHSs [3–9].

5.1 Heuristic Rules-Based Scheduling Methods

109

The performance of an AMHS is evaluated in terms of wafer production efficiency and the satisfaction degree of production strategies. The production efficiency concerns movement efficiency, throughput, use of processing tools and reduction in cycle time; production strategy issues include delivery due dates and manufacturing priorities. Several common vehicle dispatch rules are presented as follows: 1. HP rule: highest priority lot first The wafer with the highest manufacturing priority is delivered first in order to complete all of its processing steps in higher priority. 2.

ERT rule: earliest release time first The earliest released wafer is moved first.

3.

LWT rule: longest waiting time lot first The wafer with the longest waiting time is moved first in order to reduce the delivery waiting time.

4. NJF rule: nearest job first Select a recent wafer waiting to be delivered. 5.

ECD rule: earliest commit due date first The wafer with the earliest delivery date is delivered first.

6. MNJF rule: modified nearest job first It is a combination of NJF and LWT rules. The implementation is depicted in Figure 5.1. 7.

FNJF rule: factorized nearest job first It is another combination of NJF and LWT rules. Based on the above description, eight scheduling rules are tailored according to the production strategy and the manufacturing efficiency, but they cannot meet the demands of both strategy and efficiency. A fuzzy logic- based multi-mission-oriented vehicle scheduling method, which aims to improve efficiency and meet the policy requirements, is shown in Figure 5.1. The control architecture of the fuzzy logic-based multi-mission-oriented scheduler is illustrated in Figure 5.2. The fuzzy logic- based multi-missionoriented scheduler is used to determine an appropriate scheduling rule so that potentially risky wafer cards can meet the production strategy. The procedure of establishing the fuzzy scheduling model can be divided into defining input, constructing membership function, establishing fuzzy rule table and inverse fuzzification. The inputs to the fuzzy-based scheduler are based on the latency statistics and the degree to which the policy conforms to the priority and lead time requirements.

110

5 Scheduling methods of automated material handling systems in SWFSs

Yes

Is the number of available vehicles greater than the number of wafer lots waiting for moving?

Judge whether the waiting time of the longest waiting lot exceeds the predefined time factor

Execute NJF dispatching rule

No

No

Is Longest waiting time greater than the predefined time factor?

Execute NJF dispatching rule

Yes

Execute LWT dispatching rule

Figure 5.1: MNJF dispatching rule execution scenario.

Performance statistics

HP

ECD

Fuzzy-based multi-mission -oriented Inferred dispatcher rule ID Dynamic

Sampling time

switching Vehicle dispatching command

Intrabay movement model

LWT

NJF

Figure 5.2: The control architecture of the fuzzy-logic-based multi-mission-oriented vehicle dispatcher.

5.2 Operation Research Theory-Based Scheduling Methods

111

In fact, in the actual wafer manufacturing material handling process, it is difficult to use the single heuristic rule-based method to optimize the multiobjective AMHS. The weight of each parameter in the traditional compound heuristic rule is usually set to a fixed value, and it is only applicable to the static environment. How to adjust the scheduling strategy according to the changeover of system state and interference factors is a key problem to effectively schedule the material handling system in AMHSs under the dynamic and stochastic material handling environment.

5.2 Operation Research Theory-Based Scheduling Methods Operation research methods solve various systematic optimization problems by using statistics, mathematical models and algorithms so as to find the best or near optimal solution in complex problems. It has been widely applied in the field of scheduling in AMHSs.

5.2.1 Operation Research Theory The general procedure of applied operational research includes the following main steps: (1) define the nature and extent of the problem; (2) establish the mathematical model of the problem; (3) define an objective function as a scale for comparing various possible actions; (4) determine the specific value of each parameter in the model and (5) solve the model to find the optimal solution to make the objective function reach the maximum value (or minimum value). In the field of wafer material handling scheduling, operation research methods include linear programming, integer programming, dynamic programming and other methods.

5.2.1.1 Linear Programming The linear programming method is used to reasonably arrange resources under certain conditions so as to achieve the optimal economic effect. 1. Simplex method Applicable problem: all constraints are ≤; the right side are all nonnegative. There is no requirement for the coefficients of the objective function. For example, min z = 3x1 − 2x2 s.t. x1 + 2x2 ≤ 12 2x1 + x2 ≤ 18 x1 , x2 ≥ 0

112

5 Scheduling methods of automated material handling systems in SWFSs

The following steps are used to solve this problem: Step 1: Standardize the linear programming problem. Step 2: Check the initial feasible solution: if there is, go to the fourth step; otherwise, go to the third step. Step 3: Construct the auxiliary problem, and use the two-stage method to solve the auxiliary problem. If the objective function value of the optimal solution is greater than 0, there is no feasible solution and the algorithm terminates. Otherwise, go to the fourth step. Step 4: Write a simplex table, in which the coefficient of the base variable in the constraints is set to the unit matrix and the coefficient of the base variable in the objective function is set to 0. Go to the fourth step. Step 5: If the number of all non-base variables is negative or 0, the optimal solution is obtained and the algorithm terminates. Otherwise, a non-base variable whose number of tests is positive and whose absolute value is the largest is selected as the base variable. Go to the sixth step. Step 6: If the coefficients of the base variables in the constraint condition are all negative or 0, the objective function is unbounded and the algorithm terminates. Otherwise, the base variables are determined according to the minimum ratio of the right and positive coefficients. Go to the seventh step. Step 7: Transform the row of the simplex table; the principal element becomes 1 and the other elements become 0. Go to the fifth step.

2.

Dual simplex method Applicable problem: at least one of the constraints is “≥,” the corresponding right constant is nonnegative and the coefficients of the objective function are all nonnegative. The following steps are used to solve this problem: Step 1: In the first step, the initial basis B of the original problem (L) is determined, and the initial simplex table is established for all test numbers, namely, the dual feasible solution. Step 2: In the second step, the values of the base variables are checked; if ≥ 0, the optimal solution has been obtained, then terminate the search process; otherwise, L line in the simplex table corresponding to the base variable is determined as the spin-out variable. Step 3: If ≥ 0, then there is no feasible solution for the original problem, so terminate the search process; otherwise, the corresponding base variable is determined as the spin-in variable. Step 4: In the fourth step rotation transform for the new simplex table is established; go to step 2. It can be shown that the iterative solution is always the dual feasible solution.

5.2 Operation Research Theory-Based Scheduling Methods

113

5.2.1.2 Integer Planning When the variables in the plan (some or all) are limited to be integer, it is called an integer plan. The main methods include the following: 1. Cutting plane method In this method, it is possible to cut the original problem by adding new constraints in order to shrink the feasible region. The optimal solution of the original problem is gradually exposed and tends to the position of the feasible pole. It is possible to use the simplex method to find the optimal solution. The following steps are used in this method: Step 1: Solve the relaxation problem using the simplex method to obtain the optimal simplex table. Step 2: Seek a cutting plane equation, add it to the optimal simplex form and continue to solve the problem using the dual simplex method. Step 3: If integer optimal solutions are not obtained, then continue to cut the plane equation and go to the second step.

2.

Hungarian method In the real life, there are assignment problems of various natures. Assignment problems are also important issues for integer planning. For example, there are n tasks that need to be assigned to n individuals (or departments) to complete; there are n contracts need to select n bidders to contract; there are n classes need to be arranged in the classroom and so on. The basic requirements for these problems are to optimize the overall effectiveness of the assignment scheme when the specific assignment requirements are satisfied. The following steps are used in this method: Step 1: Transform the efficiency matrix so that the coefficient matrix of the assignment problem is transformed and zero elements appear in each row and column. This is done by subtracting the minimum nonzero elements of the row from the rows of the efficiency matrix, and subtracting the smallest nonzero element of the columns from the resulting coefficient matrix. The new resulting matrix will have zero elements in every row and column. Step 2: Use the circle 0 method to find the independent zero elements in the matrix C1. After the first step of transformation, the coefficient matrix of each row and each column has a separate zero element. If n is small, it can be directly observed to find n independent zero elements; while n is large, it must follow certain steps to find zero elements. (1) From a row (or column) with only one zero element, start by adding 0 to the element. This means that for the person represented by this line, there is only one task that can be assigned, and then the other elements of the column (row) are marked as ф, which means that the task represented by this column has been assigned without having to consider

114

5 Scheduling methods of automated material handling systems in SWFSs

other people. (2) Add a circle to the zero element with only one zero element column (row). (3) Repeat steps (1) and (2) until each column has no unmarked zero elements or at least two unmarked zero elements. Step 3: Test assignment. If case (1) occurs, assignment can be made: let the decision variable in the position of circle 0 be 1 and let the other decision variables be 0, and get an optimal assignment scheme and stop the calculation. In this example, after C2 is obtained, case (1) occurs. Let x14 = x22 = x31 = x43 = 1 and the remaining xij = 0 is the best assignment scheme. If case (2) occurs, then for each row, each column has two unmarked zero elements (choose one, and add the tag); that is, circle the zero element. And then mark “×” to other not marked zero elements of the same row, the same column. Case (1) or (3) may occur. If case (3) appears, go to the next step. Step 4: Make at least a straight line covering all zero elements.

5.2.1.3 Dynamic Programming Dynamic programming is a branch of algorithm design and operational research, which is an effective way to solve multi-stage optimization problems. In the middle of the twentieth century, the American mathematician Chad Berman divided the decision-making problem into multiple stages and solved the problem of each stage. He proposed the “optimization principle” to establish a large branch of operational research. Dynamic programming is the global solution for each set, and a scheme is chosen such that the objective function is minimized to find the optimal solution of the problem. Since the emergence of the dynamic programming theory, the decisionmaking process is not a linear decision-making process, it fully considers different circumstances of the decision-making process and it provides an effective way for the solution. In other words, it cannot be replaced by using an existing algorithm. According to the actual background of the given problem, a new plural is selected by using continuous decision- making process until the optimal solution is found; hence, it is called “dynamic.” Dynamic programming is established on the basis of the global optimization, to make local interest decision for each stage in order to ensure that the final global optimal is obtained. A large, complex and multidimensional variable problem is decomposed into a number of simple, small- scale variables so as to reduce the computational complexity. One of the most common shortest paths in dynamic programming is shown in Figure 5.3. In the initial stage, the optimal principle is introduced from the simple logic, the basic equation is derived on the premise of the optimal strategy and then the optimal strategy is solved by this equation. In its application process, it is found that the optimal principle is not universal for any decision-making process, and it is not unconditionally equivalent to the basic equation, and there is no definite implication

5.2 Operation Research Theory-Based Scheduling Methods

5 A

1 6 3

B1

3 B2

5

C1

D1 5

3

6

C2

D2

8

115

C3 4 C4

8 3

4

E

3 D3

Figure 5.3: Shortest path.

relationship between them. The basic equation plays a more essential role in the dynamic programming. For the initial state x1 2 X1 , the necessary and sufficient condition of the optimal strategy p*1, n = u*1 , ..., u*n is for any k, 1 ≤ k ≤ n, there is    V1, n x1 , p*1, n = ’ optp1, k − 1 2p1, k − 1 ðx1 Þ ½V1, k − 1 ðx1 , p1, k − 1 Þ, optp

½Vk, n ðxm , pk, n Þ k, n 2pk, n ðxk Þ



(5:1) If p*1, n 2 p1, n ðx1 Þ is the best solution, for any k (1 τij ðtÞ  ηij ðtÞ > > < iβ

h pkij ðtÞ = P τij ðtÞ α  η ðtÞ ij > > > > s ∉ Γk : 0

if j ∉ Γk

(5:14)

if j 2 Γk

where α is the information heuristic factor that indicates the relative importance of pheromone accumulated in the movement process and changes the dependence of ant on the existing pheromone. The larger the α value, the greater the impact of path information on the decision-making of ants and the more the probability selection consists of the collaboration among ant colony. β is the expectation heuristic factor that represents the path selection and the dependence on visibility (such as path distance). The larger the β value, the greater the impact of the path distance on ant selection and the more the probability selection consists of the greedy rule. ηij(t) is the value of the heuristic function, which is calculated by using the following formula: ηij ðtÞ =

1 dij

(5:15)

In formula (5.15), dij denotes the distance lij between two adjacent nodes. For ant k, the smaller the dij, the larger the ηij(t). Excessive residual pheromone trail on certain paths directly dispatches the probability selection so as to cause undesirable effects. In order to avoid excessive residual pheromone, the residual pheromone is globally updated after each ant has visited all the nodes (n nodes). At the moment t + n, the pheromone on the path ði, jÞ is updated according to formula (5.16): τij ðt + nÞ = ð1 − ρÞ  τij ðtÞ + Δτij ðtÞ

Δτij ðtÞ =

m X k=1

Δτkij ðtÞ

(5:16)

(5:17)

where ρ is the pheromone evaporating rate and 1 – ρ is the residual pheromone factor. In order to prevent unlimited accumulation of pheromone, ρ is to be ρ  ½0, 1. Δτij(t) is the pheromone increment on the path ði, jÞ in this cycle, and Δτij(0) = 0 at the initial time. Δτkij (t) is the pheromone increment of ant k on the path ði, jÞ in this cycle. According to the pheromone update strategy, Dorigo proposed three different basic ant colony models, respectively, called Ant-Cycle model, Ant-Quantity model and Ant-Density model. Their difference lies in how to achieve Δτkij (t).

5.3 Artificial Intelligence-Based Scheduling Methods

127

(a) In Ant-Cycle model: 8 |V|), while this case is called task-prioritized dispatching; (2) the number of waiting cassettes is smaller than the number of idle vehicles (|R| ≤ |V|), while this case is called vehicle-prioritized dispatching; and (3) the buffer of stocker is limited in the system, the vehicles have to wait until the stocker has enough storage space and this case is called buffer-prioritized dispatching.

6.2 AMPHI-Based Interbay Automated Material Handling Scheduling Methods Although the existing research works on Interbay scheduling problems which focused on heuristic rules have shown good efficiency and usability, it is hard to implement them to meet multi-objective optimization. Some scheduling methods based on intelligent algorithms or mathematical programming can describe the optimization target more accurately and can realize multi-objective optimization; however, it is difficult to implement them in a real-time way. In this section, an AMPHI (adaptive multi-parameter and hybrid intelligence)-based Interbay material scheduling method is proposed to implement production scheduling in a real-time way by combining a modified Hungarian algorithm (MHA) and a fuzzy-logic-based controlling method [1, 2, 3].

6.2.1 AMPHI-Based Interbay Automated Material Handling Scheduling Model A mathematical model is established to describe the characteristics of an Interbay material handling system, which includes the performance optimization indicators related to normalized wafer lots’ delivery time, wafer lots’ total delivery time, transport amount, vehicle utilization, wafers’ throughput, cycle time, WIP quantity and wafer’s delivery due date satisfactory rate. 1. Notations Tu – Smallest time unit in the material handling process (vehicle speed, loading/ unloading time and wafer delivery request arrival time are integer multiples of Tu); T – Cycle time period in the Interbay material handling system, T 2 fTu , 2Tu , 3Tu , ...g; Y – Set of tracks and shortcuts in the Interbay material handling system; S – Set of stockers in the Interbay material handling system; etyj – etyj = 1 denotes that vehicle j is located at track or shortcut y at time t, y 2 f1, 2, ..., jY jg; αu – Weighted coefficient of system performance indexes, αu 2 ð0, 1, u = 1, ..., 8; PDs – Set of pick-up/drop-off events at stocker s, PDs = f..., ðk, iÞ, ...g, s 2 S; PrðiÞ – Product type of wafer i;

138

6 Scheduling in Interbay automated material handling systems

DeðiÞ – Position of destination stocker or processing machine of wafer i; ReðiÞ – Position of stocker or processing machine of waiting wafer i; w – Within cycle time T, the quantity of wafers waiting for transportation in the Interbay system; n – Within cycle time T, the quantity of wafers delivered by vehicles; m – The quantity of vehicles in the Interbay system; d – Maximum time in which a vehicle can transport a wafer lot; Tmax t Tmax – Maximum value of wafer lots’ total transportation time; C Tmax – Maximum value of wafer lots’ total cycle times; Lmax – Maximum value of wafer lots’ throughput; twi – The moment that wafer lot i sends out a transportation request; tpi – The moment that wafer lot i is loaded; tdi – The moment that wafer lot i is unloaded; h – The time that the vehicle spends in loading/unloading a wafer lot; tri – The moment that wafer lot i is released to the Interbay system, tri subject to certain probability distribution; i – The moment that wafer lot i leaves the Interbay system after finishing all tout processing steps; q – Within T, the quantity of wafer lots entering the Interbay system; p – Within T, the quantity of wafer lots leaving the Interbay system; l – Within T, the quantity of wafer lots leaving the Interbay system after finishing all processing steps; F i – Due date relaxation coefficient of wafer lot i; Tpi – Total processing time of wafer lot i; TDi – The shortest delivery time for wafer lot i to move from the current stocker to the destination stocker; Jw – Set of water lots at stocker w; 2.

Decision variables Xijt – A 0–1 decision variable where Xijt is equal to 1, if wafer lot i is transported by vehicle j at moment t; otherwise, Xijt is equal to 0.

3.

Objective function 0

n n n P P P ðtdi − tpi + 2hÞ ðti − ti Þ ðtdi − twi Þ B i=1 d p ðw − nÞ i = 1 i = 1 Obj = MinB + α4 + α2 + α3 @α1 n × T d m×T n × Tt w max

max

l P

+ α5

v ðtout − trv Þ

ðLmax − lÞ + α6 v = 1 C l*Tmax Lmax

1 (6:1) l v ðq − pÞ α8 X ðtout − trv ÞC C + + α7 C v l v = 1 F × Tpv A q

6.2 AMPHI-Based Interbay Automated Material Handling Scheduling Methods

139

4. Constraints m X

Xijt = 1, Xijt 2 f0, 1g, ∀t 2 ½0, T, i = 1, ..., n

(6:2)

j=1 n X

Xijt = 1, Xijt 2 f0, 1g, ∀t 2 ½0, T, j = 1, ..., m

(6:3)

i=1 tk

ti

tdk Xkjd + h + 1 ≤ tpi Xijp ti

ti

tpi Xijp + TDi + h ≤ tdi Xijd ti

tk

(6:4) (6:5)

tpk Xkjp + h ≤ tdi Xird , ∀ ReðiÞ = DeðkÞ, ðk, iÞ 2 PDs

(6:6)

tri < twi < tpi

(6:7)

tpk + h ≤ tpi , ∀ i, k 2 Jw : ðtrk > tri Þ

(6:8)

m X

etyj ≤ 1, ∀t 2 ½0, T, ∀ y 2 Y

(6:9)

j = 1‘

where formula (6.1) is the objective function to minimize the weighted summation of normalized wafer lots’ delivery time, wafer lots’ total delivery time, transport amount, vehicle utilization, wafers’ throughput, cycle time, WIP quantity and wafer’s due date satisfactory rate; formula (6.2) ensures that each wafer lot can be transported by only one vehicle; formula (6.3) makes sure that a vehicle can transport a wafer lot only at a time; formula (6.4) ensures that the vehicle cannot start next transportation task unless it has completed current task; formula (6.5) makes sure that the loading and unloading times of the same wafer lots must satisfy certain constraints; formula (6.6) ensures that the loading and unloading times of a wafer lot at a stocker must satisfy certain constraints; formula (6.7) makes sure that the times that a wafer lot enters the Interbay material handling systems, makes transportation request and is transported by a vehicle should meet certain time constraints; formula (6.8) ensures that loading times of different wafer lots in the same stocker must meet chronological constraints and formula (6.9) ensures that each track, shortcut or turntable can only carry one vehicle at any time.

6.2.2 Architecture of AMPHI Interbay The architecture of the proposed AMPHI-based Interbay automated material handling scheduling system is shown in Figure 6.3. First, multiple parameters including

140

6 Scheduling in Interbay automated material handling systems

Work load balance factor System’s load ratio factor System’s waiting time factor Lots’due date factor

Takagi-Sugeno fuzzy logic-based multiparameter weight modification model

Vehicle’s distance parameter (VDP) Lot’s due date parameter (DDP) Lot’s waiting time parameter (WTP) Lot’s origin-destination buffer status parameter (OBS)

Multi-parameter’s weight coefficient 󰡷1, 󰡷2, 󰡷3, 󰡷4

Vehicle dispatching model of Interbay material handling system

Cij = 󰡷1 × VDPij + 󰡷2 × DDPi + 󰡷3 × (1—WTPi ) + 󰡷4 × OBSi

Hungarian algorithmbased vehicle dispatching policy

c11 ... c1m C1 = ... cij ... cn1 ... cnm

Automated vehicle dispatching Interbay material handling system in a semiconductor wafer fabrication fab

Figure 6.3: Architecture of the AMPHI-based Interbay AMHS.

Vehicle’s transportation distance, lot’s due date, lot’s waiting time and lots’ origindestination buffer status are considered to develop a multi-parameter wafer handling cost model in order to meet the multi-objective scheduling requirements of an Interbay and balance the WIP in each processing district. Second, to reduce the temporary blockage of an Interbay material handling system, waiting wafer sets are optimally matched to empty vehicles in order to obtain a real-time material handling assignment scheme. Finally, according to the dynamic environment of an Interbay material handling system, a weight modification model is proposed based on fuzzy logic theory to adaptively adjust the weight values of these model parameters in order to optimize the multi-objective Interbay scheduling problem.

6.2.3 Vehicle Dispatching in the Interbay Material Handling System A mathematical model is developed to solve scheduling problems in the Interbay material handling system. Suppose at time t when the cassette arrives at the Interbay material handling system or the vehicle becomes idle, there are n waiting cassettes and m available vehicles. This scheduling problem in the Interbay material handling system is formulated as follows:

Min

n X m X i=1 j=1

! cij Xij = MinðC  XÞ

(6:10)

6.2 AMPHI-Based Interbay Automated Material Handling Scheduling Methods

n X

Xij = 1,

i=1

m X

Xij = 1

cc11 C=4 ... cn1

(6:11)

j=1

Xij 2 f0, 1g, i = 1, 2, . . . , n, j = 1, 2, . . . , m 2

141

... cij ...

3 c1m ... 5 cnm

(6:12)

(6:13)

where X is the matrix of matching factor. Xij = 1 indicates that the vehicle j is assigned to deliver the cassette i; otherwise Xij = 0. C is the matrix of delivery cost. cij is the cost value of assigning vehicle j to cassette i. 1. Hungarian algorithm-based vehicle dispatching policy In the Interbay material handling system, multiple parameters including Vehicle’s transportation distance lot’s due date, lot’s waiting time and lots’ origin-destination buffer status are taken into consideration, and cij is calculated by the below-weighted multi-parameter wafer handling cost function: cij = w1  VDPij + w2  DDPi + w3  ð1 − WTPi Þ + w4  OBSi

(6:14)

In eq. (6.14), cij is calculated as the weighted sum of parameters for vehicle j’s distance to lot i (VDPij), lot i’s due date (DDPi), lot i’s waiting time (WTPi) and lot i’s origin-destination buffer status (OBSi ). w1, w2, w3 and w4 are weight coefficients for parameters VDPij, DDPi, WTPi and OBSi, respectively. w1, w2, w3 and w4 are adaptively adjusted by the Takagi-Sugeno fuzzy-based multi-parameter weight modification model. Multiple variables including Vehicle’s transportation distance and waiting time are considered to assign vehicle, and in order to integrate these different variables, they are first normalized as ratio (parameters) by dividing itself to the maximum possible value. For example, VDPij denotes the normalized travel by dividing the distance between the current position of vehicle j to the position of job i to the maximum travel distance of vehicle. Similarly, WTPi denotes the normalized waiting time parameter for job i. DDPi denotes the due date parameter of candidate job i. DDPi can be calculated by DDPi = RTi/(RPi × Facmax), where RTi is the remaining time before the due date of wafer lot i, RPi is the remaining processing time of wafer lot i and Facmax is the maximum due date slack coefficient of completed wafer lots. OBSi denotes the origin-destination buffer status parameter of job i, which is transported between i i i × (1–Bout ), where Bin is the two stockers. OBSi can be calculated by OBSi = Bin i normalized buffer level of the input buffer in the destination stocker and Bout is the normalized buffer level of the output buffer in the source stocker.

142

6 Scheduling in Interbay automated material handling systems

The solution procedure using the Hungarian method is described as follows. Step 1. According to the weighted multi-parameter wafer handling cost function (eq. (6.14)), a cost matrix C1 is first constructed. Step 2. For the original cost matrix C1, if the number of available vehicles is more than that of candidate jobs, that is, m > n, (m × n) dummy candidate jobs are added to form an m × m cost matrix C2. Otherwise, if the number of candidate jobs is more than that of available vehicles, that is, m × n, (n × m) dummy vehicles are added to form an n × n cost matrix C2. All dummy cost values are cij = max{cij}. Let k denote the order of cost matrix C2, k = max(m, n). Step 3. A reduced cost matrix C3 is obtained by subtracting from all elements in each row of C2, and the minimum element of that row is c1j (j = 1,. . .,k). Step 4. Another reduced cost matrix C4 is obtained by subtracting from all elements in each column of C3, and the minimum element of that column is ci1 (i = 1,. . .,k). Step 5. Cross out rows or columns with element cij = 0 in C4. Let km be the minimum number of lines needed to cross out all zero elements in C4. If km = k, then the optimal dispatching solution is found and the procedure is stopped. Otherwise, go to Step 6. Step 6. Let d be the minimum element in the remaining (not crossed out) elements of C4. A new reduced cost matrix C4 is obtained by subtracting d from all remaining elements in C4 and adding d to elements being double crossed out (elements at the intersection of crossed-out row and column) in C4. Go to Step 5. 2.

Modified Hungarian algorithm-based vehicle dispatching policy In MHA-based vehicle dispatching policy, four major parameters are taken into account when calculating the cost value of assigning a vehicle to a wafer cassette: delivery time TP_factor, cassette waiting time WT_factor, cassette due date Due_factor and cassette processing factor PF_factor. Suppose the vehicle j is designated to deliver cassette i, the cost value cij is defined as follows: cij = w1 × TP factorði, jÞ + w2 × WT factorðiÞ + w3 × Due factorðiÞ + w4 × PF factorði, n, cÞ

(6:15)

where TP_factor(i, j) is the linear normalized delivery time of assigning vehicle j to cassette i, and TP factorði, jÞ = disði, jÞ=ðarg maxðdisði, jÞÞÞ. i2L, j2K

WT_ factor(i) is calculated from linear normalization of cassette i’s waiting i time in the Interbay material handling system, WT factorðiÞ = ðWTmax − WTi Þ= i is the aging time for cassette i and WTi is the time period for WTmax :WTmax which cassette i has been waiting in the Interbay.

6.2 AMPHI-Based Interbay Automated Material Handling Scheduling Methods

143

Due_factor denotes the due date parameter of candidate job i, and ðDUEi − tRi Þ ðDUEi − tRi Þ Due factorðiÞ = = arg max , DUEi is the due date of ðAPTi − PTi Þ ðAPTi − PTi Þ i2L waiting cassette i.tRi is the remaining time to the due date of cassette i.APTi is the sum of processing time of all operations. PTi is the sum of processing time of all finished operations. Smaller Due_factor indicates that the job is more urgent. PF_factor is designed to measure the buffer level of source and destination stockers for cassette i, and PF factorði, n, cÞ = ðQmax + Qic − Qin Þ=2Qmax , Qmax is the maximum of Qin and Qic . Qin denotes the queue length in the in-buffer of destination stocker n while Qic denotes the queue length in the out-buffer of source stocker c. In eq. (6.15), w1, w2, w3 and w4 are weight coefficients for parameters TP_factor (i, j), WT_factor(i), Due_factor(i) and FF_factor (i, n, c), respectively. w1, w2, w3 and w4 are adaptively adjusted by the Mamdani-fuzzy-logic-based multi-parameter weight adjusting method. The Hungarian algorithm was first proposed by the Hungarian mathematician Egervary and then improved by Edmonds on the basis of the Berge Theorem and Hall’s Theorem. The basic Hungarian method is a combinational optimization algorithm, which effectively optimizes solutions for assignment problems. It can be used to solve the maximal matching problem that whether there exists saturated X or Y in a bipartite graph G = (X, Y). As a stable polynomial-time algorithm, it ensures to get the optimal solution, in which when |X | = n and |Y | = m, the time complexity of the algorithm is Oðv3 Þ. The MHA is proposed to solve the maximal matching problem for a weighted bipartite graph. It is essentially a labeling method with the theoretical basis of feasible vertex labeling method. In detail, suppose in a weighted bipartite graph G = (X, Y), X is the set of cassettes and Y is the set of vehicles. ∀i  X and ∀j  Y, each vertex is given a label denoted by l(i) and l(j). For any edge, e = ði, jÞ, if lðiÞ − lðjÞ ≤ wði, jÞ = cij

(6:16)

this labeling is called a feasible vertex labeling. An edge can be defined as el = fði, jÞ 2 EðGÞjlðiÞ − lðjÞ = wði, jÞg

(6:17)

in which E(G) denotes the set of all edges in a weighted bipartite graph G = (X, Y). The sub-graph with el as its edges is called the equal-sub-graph Gl which is depicted in Figure 6.4. According to the Berge Theorem and Hall’s Theorem, if there exists a perfect match M* in the equal-sub-graph Gl (each vertex in Gl is related to M* ), then M* is the minimum weight perfect match. According to the above mentioned theoretical basis, the solution procedure using the proposed MHA is described as follows.

144

6 Scheduling in Interbay automated material handling systems

l(x1)

y1

y2

yn

x1 c11

c12

c1n

x2 c21

c22 cnn

xn cv1 l(x1) x1

l(x2) x2

l(x2)

x1

y1

xn

y2

l(y1) l(xn) xn

l(xn)

x2

yn

l(y2)

l(yn)

l(x1) – l(y2) = c12 l(x2) – l(y1) = c21 l(x2) – l(y2) = c22 l(x2) – l(yn) = c2n

y1 l(y1)

y2 l(y2)

Figure 6.4: The relationship diagram of feasible vertex labeling and equal subgraph.

yn l(xn) – l(y2) = cn2 l(yn)

Step 1. The initial feasible vertex labels are given by eq. (6.18). Based on these labels, establish the equal-sub-graph Gl . Search the perfect match M* in Gl . α indicates one of the waiting cassettes and β indicates one of the vehicles. (

min ðwðα, βÞÞ, α 2 X

∀β2Y

lðαÞ =

α2Y

0

(6:18)

Step 2. If there exists a perfect match M* in Gl , then M* is the minimum weight perfect match and the algorithm terminates. Otherwise, the Hungarian algorithm terminates at S  X, T  Y and NGl ðSÞ = T. S and T denote the set of matched X(cassettes) and Y(vehicles). NGl ðSÞ stands for the set of neighbor vertices for S in Gl. Step 3. Let SlackðlÞ = minfwði, jÞ − lðiÞ + lðjÞji 2 X, j 2 Y − Tg, ∀l 2 Y, for any j 2 Y − T, the vertex labels are modified tol′ðjÞ = lðjÞ + SlackðlÞ; while the others are fixed. Such labeling is still feasible and at least there exists one edge which is included into a new equal-sub-graph Gl′ , which indicates that Gl ′ is scalable. Step 4. Repeat Steps 2 and 3 until an equal-sub-graph Gl ′ with a perfect match is obtained. Denote the weight matrix of Gl ′ as Cr. Step 5. If there is more than one perfect match, this means there exists more than one optimal solution. Further procedures are as follows: (a) Let σ2ij =

ðCij − μÞ2

, μ=

m P n P

Cij , Cmax and Cmin denote the maximal and ðCmax − Cmin Þ i=1 j=1 minimal elements in the initial cost matrix C, and σ2ij is the normalized 2

variance; (b) Modify the initial cost matrix. For those elements in which Cr is valued 0, add σ2ij to them; for those non-zero elements, reset them to Cmax . The new cost matrix CNew is constructed as

6.2 AMPHI-Based Interbay Automated Material Handling Scheduling Methods

0

Co11 + σ211 B Cmax B CNew = B .. @ . Cmax

Cmax Co22 + σ222 .. .

Con2 + σ2n2

  .. . 

Co1n + σ21n Cmax .. .

145

1 C C C A

(6:19)

Conn + σ2nn

Go to Step 1. When there are multiple optimal matches, the above method can be used to obtain a solution with smaller variance, and the whole process is convergent.

6.2.4 Fuzzy-Logic-Based Weight Adjustment When assigning a vehicle to a wafer cassette, we must first determine the weight relationship between the four parameters in the wafer handling cost model, and improper parameter weighting will decrease the actual operation efficiency. If the weight coefficient wj is set to a fixed value, the Interbay material handling task assignment method mentioned in section 6.2.3 applies only to the material handling system in the static environment. This section describes fuzzy-logic-based weight adjustment methods, which dynamically adjust arrangement of Interbay material handling tasks according to the system status and interference factors in dynamic stochastic environments. Fuzzy logic inference is in essence a process that maps a given input space to a specific output space by using a fuzzy logic method. According to fuzzy input and fuzzy rules, fuzzy outputs are obtained by using good reasoning methods. There are three common fuzzy inference methods: pure fuzzy logic reasoning, Mamdani fuzzy inference and Takagi-Sugeno fuzzy inference. This section introduces two methods to adjust parameters of wafer handling cost model: Mamdani fuzzy inference and Takagi-Sugeno fuzzy inference. 1. Takagi-Sugeno fuzzy-based weight modification model In this section, a weight modification model is developed based on the Takagi-Sugeno fuzzy approach to define the weight coefficients in cost function of eq. (6.10). The main procedures for developing such fuzzy-based weight modification model include (1) defining input variables, (2) constructing membership functions and fuzzy rule set and (3) deploying TakagiSugeno’s fuzzy inference. The principle of the proposed fuzzy-based model is to dynamically adjust the weights to be put on different criteria for assigning vehicle to individual jobs based on the overall system loading and job completion status. Therefore, four fuzzy input variables, namely, the Interbay system’s load ratio factor, the lots’ due date factor, the system’s waiting time factor and the system’s workload balance factor, are defined to represent the overall system dynamic status.

146

6 Scheduling in Interbay automated material handling systems

1)

Input variables (1) System’s load ratio factor (LRF). The LRF is used to measure the current transportation load of the Interbay material handling system. A high LRF score indicates that the Interbay system is heavily loaded. The LRF is calculated as LRF =

n1 X

WSs =NV

(6:20)

s=1

where WSs is the number of wafer lots waiting in the output buffer of stocker s, N is the total number of stockers and NV is the total number of automated vehicles in the Interbay system. (2) Lots’ due date factor (DDF). The DDF is used to measure the current wafer lots’ due date satisfaction. A larger DDF value indicates a higher probability of satisfying the committed due date. DDF is calculated as

DDF =

m1 X i=1

RTi 1 × RPi × Fac m1

(6:21)

where RTi is the remaining time before the due date of wafer lot i, RTi is the remaining processing time of wafer lot i, m1 is the total number of waiting wafer lots in the Interbay system and Fac is the due date slack coefficient of completed wafer lots. Fac can be calculated as follows: h CT P 1 q × , where PTq is the overall processing time of wafer lot q, Fac = PT h q q=1 CTq is the cycle time of wafer lot q and h is the number of wafer lots that have completed all processes in an SWFS. (3) System’s waiting time factor (WTF). The WTF is used to measure the lateness of wafer lots in the Interbay system. A higher WTF value indicates that more wafer lots have a longer than expected waiting time. WTF is calculated by WTF = Nwt=m1

(6:22)

where Nwt is the number of wafer lots whose waiting time is greater than the expected average hwaiting time AW in the Interbay system. P CWq =h, where CWq is the mean waiting AW is calculated as AW = q=1 time of wafer lot q. (4) System’s workload balance factor (WBF). The WBF is used to measure the starvation or blockage probability of the Interbay system. A higher WBF value means a higher chance that the stockers in the Interbay system are blocked or starved. WBF is calculated by

147

6.2 AMPHI-Based Interbay Automated Material Handling Scheduling Methods

1 1X  2 ðBs − BÞ n1 s = 1

n

WBF =

(6:23)

 is the where Bs = expð1 − ðBSs =NsÞ2 Þ is the load factor of the stocker buffers, B mean load factor of the Interbay system, BSs is the buffer level of stocker buffers and Ns is the buffer capacity of stock buffers. 2)

Membership function and fuzzy rule table The membership functions of the above-defined input variables are represented in the form of triangular and trapezoidal fuzzy sets in the proposed fuzzy-based weight modification model. Table 6.1 lists the membership functions of the four fuzzy input variables. Let βw1 , βw2 , βw3 and βw4 be the exact values of fuzzy output variables w1 , w2 , w3 andw4 in the fuzzy rule. By Takagi-Sugeno fuzzy inference, the weight coefficient wr (r = 1,…,4) of eq. (6.10) can be evaluated and the fuzzy rule table is established, as shown in Table 6.2. Both Tables 6.1 and 6.2 are obtained by consulting veterans in the field of AMHSs and SWFSs, and conducting extensive simulations based on historical production data using design of experiments (DOE).

3) Takagi-Sugeno’s fuzzy inference The proposed fuzzy-based weight modification model adjusts the weight coefficients w1 , w2 , w3 and w4 in the cost function (eq. 6.10) by evaluating the input variables using a set of fuzzy rules. The fuzzy rule has the following generic form:

Table 6.1: The membership function of the input variables of weight modification model. Input variables

Fuzzy sets (linguistic descriptions)

fuzzy set membership functions

LRF

Low load (LL) Medium load (ML) High load (HL) Urgent due date (UDD) Normal due date (NDD) Slack due date (SDD) Short waiting time (SWT) Normal waiting time (NWT) Long waiting time (LWT) Good balance (GB) Normal balance (NB) Bad balance (BB)

( – ∞, 0.0, 0.7, 1.0) (0.7, 1.0, 1.3) (1.0, 1.3, 5.0, + ∞) ( – ∞, 0.0, 0.8, 1.0) (0.8, 1.0, 1.5) (1.0, 1.5, 6.0, + ∞) ( – ∞, 0.0, 0.2, 0.35) (0.2, 0.35, 0.5) (0.35, 0.5, 1.0, + ∞) ( – ∞, 0.0, 0.2, 0.35) (0.2, 0.35, 0.5) (0.35, 0.5, 1.0, + ∞)

DDF

WTF

WBF

148

6 Scheduling in Interbay automated material handling systems

Table 6.2: Fuzzy rule table of the multi-parameter weight modification model. Fuzzy rule ID

FR1 FR2 FR3 FR4 FR5 FR6 FR7 FR8 FR9 FR10 FR11 FR12 FR13 FR14 FR15

Input linguistic variables

Crisp output variables

LRF

DDF

WTF

WBF

HL HL HL HL HL HL ML ML ML ML ML ML LL LL LL

UDD UDD UDD / / / UDD UDD UDD / / / / / /

/ / / LWT LWT LWT / / / LWT LWT LWT / / /

GB NB BB GB NB BB GB NB BB GB NB BB GB NB BB

βw1

βw2

βw3

βw4

49 33 25 49 33 25 49 25 1 49 25 1 97 49 1

49 33 25 1 1 1 49 49 49 1 1 1 1 1 1

1 1 1 49 33 25 1 1 1 49 49 49 1 1 1

1 33 49 1 33 49 1 25 49 1 25 49 1 49 97

~ j AND DDFj is Q ~ j AND WTFj is Q ~ j AND WBFj is Q ~ jm , THEN IF LRFj is Q 1 2 3 j βjwr = ajr (r = 1, 2, 3, 4) ~ j (i = 1,…, 4, j = 1, 2,…, Nr) are fuzzy sets; LRFj, DDFj, WTFj and WBFj where Q i are input variables; j is the rule number and Nr is the total number of fuzzy rules. βjwr ðr = 1, . . . , 4Þ is the output variables of the jth fuzzy rule and ajr ðr = 1, . . . , 4Þ is the output coefficient of the jth fuzzy rule. By TakagiSugeno fuzzy inference, the weight coefficient wr (r = 1,…, 4) of eq. (6.10) can be evaluated as a weighted mean of βjwr for all fuzzy rules. 15 P

wr =

j=1

vj × βjwr 15 P

, r = 1, ..., 4

(6:24)

vj

j=1

where the weight vj is the degree of relevance of the premises of the jth fuzzy rule for input variables LRF, DDF, WTF and WBF. In the Takagi-Sugeno fuzzy inference, vj is calculated as vj = μQ~ j ðLRFÞ × μQ~ j ðDDFÞ × μQ~ j ðWTFÞ × μQ~ j ðWBFÞ 1

2

3

(6:25)

4

where μQ~ j ðÞ are the membership functions of input variables LRF, DDF, WTF and WBF.i

149

6.2 AMPHI-Based Interbay Automated Material Handling Scheduling Methods

Mamdani fuzzy rules i 1

i i i i R 1 : If x is A 1i and … xn is An Then ym = Bm … … … … …

n

n

R n: If x1 is A1n and … xn is Ann Then ymn = Bmn Input: Interbay real-time parameters

(RTP, POD, DSF, LBF)

Weight

Input fuzzy

Fuzzy inference

Defuzzification

w =[󰡷1, 󰡷2, 󰡷3, 󰡷4 ]

Figure 6.5: The framework of fuzzy logic-based weight-adjusting method.

2.

Mamdani fuzzy-based weight modification model In this section, a Mamdani fuzzy-logic-based method is proposed to adjust the weight coefficients in eq. (6.15). The framework of this fuzzy-logic-based weightadjusting method is shown in Figure 6.5. The main procedures include (1) define input variables, (2) determine the domain of discourse and membership function, (3) construct a fuzzy rule set and (4) synthesization and defuzzification. (1) Input and output variables (1) Ratio of transport and processing load (RTP) The RTP is used to measure the current transportation load of the Interbay material handling system. RTP is calculated as

RTP =

m X n X

expðð1=TDij Þ RÞ

(6:26)

j=1 i=1

Among the above equations, n is the total number of waiting wafer cassettes in the current Interbay material handling system. m is the total number of available vehicles. TDij denotes the expected shortest time for which vehicle j arrives at the stocker where the cassette i locates. R is the adjustment factor which is the mean ratio of the actual delivery time for all finished job DTi and the next operation processing time IPi, l′ðjÞ = lðjÞ + SlackðlÞ. A high RTP indicates that the Interbay material handling system is heavily loaded, and the mean delivery time is too long. (2) Due date satisfaction rate (DSF) The DSF is used to measure the current wafer lots’ due date satisfaction. DSF is calculated as follows:

DSF =

n X i=1

RTi 1 × RPi × Fac n

(6:27)

150

6 Scheduling in Interbay automated material handling systems

Among the above equations, RTi = DUEi – t is the remaining time before the due date of wafer lot i at time t; RPi = APi – PTi is the sum of processing time for remaining operations. Fac is the adjustment factor which is the ratio of the actual processing cycle time for finished wafer lots and the total processing time for all operations. Gl ′. A larger DSF value indicates a higher probability that the lots cannot meet the committed due date. (3) Percent of delay (POD) The POD is used to measure the lateness of wafer lots in the Interbay system. POD is calculated as POD =

1 jfij i 2 M, WTi > AW gj n

(6:28)

Among the above equations, AW is the mean waiting time for all waiting n P wafer lots. AW = CWi × n1 . CWi is the waiting time of cassette i. A i=1

higher POD value indicates that more wafer lots have a longer than expected waiting time. (4) Load balance factor (LBF) The LBF is used to measure the balance of stockers in the Interbay material handling system. LBF is calculated as LBF =

n 1 X  2 ðBi − BÞ n − 1 i=1

(6:29)

Among the above equations, Bi is the load factor of the buffer in Stocker i.   BS n P − λ Capi − c0 1  B = n Bi , Bi = 1=ð1 + e Þ. BSi is the number of waiting casi=1

settes. Bi is a bell-shaped curve function, and λ and c are shape control parameters. A higher LBF value indicates the imbalance of current system load. The output of the proposed fuzzy logic control (FLC) is w = [w1, w2, w3, w4]. w1, w2, w3 and w4 are weight coefficients for the proposed four input variables. 2)

Domain of discourse and membership function The domains of discourse for the input variables are determined by orthogonal design test method. In each case, simulations are carried out for different combinations of discourse domains. The best domains are obtained as URTP = [0.5, 1.5], UDSF = [1.5, 2.5], UPOD= [0.3, 0.7] and ULBF = [0.6, 1.4]. For each input variable, there are three fuzzy sets (Table 6.3).

6.2 AMPHI-Based Interbay Automated Material Handling Scheduling Methods

151

Table 6.3: The fuzzy set of the input variables of FLC.

Low rate (LR) Slack due date (LDS) Short waiting time (SD) Poor balance (NLB)

M(POD)

Medium load (MR) Normal due date (MDS) Normal waiting time (MD) Normal balance (MLB)

1 HR MR 0.9 LR 0.8 gaussmf (0.15,1.0) zmf (0.5,1.0) smf (1.0,1.5) 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.6 0.8 1 1.2 1.4 1.6 0.4 RTP 1 SD LD MD 0.9 0.8 zmf (0.3,0.5) gaussmf (0.05,0.5) smf (0.5,0.7) 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.2 0.3 0.4 0.5 0.6 0.7 0.8 POD

M(DSF)

RTP DSF POD LBF

M(LBF)

Fuzzy set

M(RTP)

Input variable

High load (HR) Urgent due date(HDS) Long waiting time (LD) Good balance (GLB)

1 HDS MDS LDS 0.9 0.8 zmf (1.5,2.0) gaussmf (0.2, 2.0) smf (2.0,2.5) 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 DSF 1 GLB NLB MLB 0.9 0.8 zmf (0.6,1.0) gaussmf (0.15,1.0) smf (1.0,1.4) 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.4 0.6 0.8 1 1.2 1.4 1.6 LBF

Figure 6.6: The membership–function curves of the input variables of FLC.

The membership functions of the above-defined input variables are represented in the form of Gaussian function, Z-function and S-function in the proposed fuzzy-logic-based weight adjusting method, as shown in Figure 6.6. The domain of discourse for the output wi(i = 1, 2,…, 4) is U = [0, 1]; the membership function is shown is Figure 6.7. 3) Fuzzy rules The Mamdani-based weight adjusting method modifies the weight coefficients w1, w2, w3 and w4 in the cost function eq. (6.15) by evaluating the

152

6 Scheduling in Interbay automated material handling systems

M(wi) 1.0

S

M

0.2

0.5

L

0 0.7 0.8

1.0 wi

Figure 6.7: The membership–function curve of the output parameter of FLC.

Table 6.4: The fuzzy rules table of the Mamdani-based inference method. Rule

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

Input linguistic variables

Output variables

RTP

DSF

POW

LBF

α1

α2

α3

α4

HR HR HR HR HR HR HR HR HR MR MR MR MR MR MR MR MR MR LR LR LR LR LR

LDS MDS HDS / / / / / / LDS MDS HDS / / / / / / LDS MDS HDS / /

/ / / LD MD SD / / / / / / LD MD SD / / / / / / LD /

/ / / / / / GLB NLB BLB / / / / / / GLB NLB BLB / / / / BLB

L L L L L L M M M M M M M M M M M M S S S S S

L M S S S S L M S S S S L S S S S S L M S S S

S L S L M S S S S L M S L M S S S S S S S L S

S L S S S S S S S S S S S S S S M L S S S S L

input variables using a set of fuzzy rules. The output of the Mamdani inference is Zi = ½yi1 , yi2 , yi3 , yi4 , i = 1, 2,…, nr, with nr being the total number of fuzzy rules. The fuzzy rule has the following generic form:

6.2 AMPHI-Based Interbay Automated Material Handling Scheduling Methods

153

~ i 1 and...xi is A ~ i n then yi = B ~ i ðm = 1, 2, 3, 4Þ If xi 1 is A m ni m i fi ðk = 1, 2, ..., n Þ is the i xk ðk = 1, 2, ..., ni Þ is the kth input variable of rule i. A i k f corresponding fuzzy set for the input variable. Bi is the fuzzy set for the mth k

output variable yim of rule i. The fuzzy implication of A→B is calculated as ð Rc = A ! B =

μA ðxÞ ^ μB ðyÞ ðx, yÞ

(6:30)

X×Y

There are four input variables and three fuzzy sets for each variable. Twentythree among those 81 combinations are chosen. Table 6.4 lists the fuzzy rules adopted in the proposed weight adjusting method. 4) Synthesization and defuzzification Mamdani inference adopts the Max method to carry out the synthesization for all output fuzzy sets in coincidence with all fuzzy rules. The final output is C′ = C1′ ∪ C2′ ∪ ... ∪ Cn′ r = ∪ ni =r 1 Ci′

(6:31)

Ci ′ is the output fuzzy set for rule i. The centroid method is applied to realize the defuzzification. Suppose Cm ′ is the final output fuzzy set of variable wm on its discourse domain Um; then wm is Ð wm =

Um

Ð

j=Δx jUm P

μCm′ ðxÞ.xdx

Um

μCm′ ðxÞdx

=_

k=1 jUm j=Δx P k=1

μCm′ ðxk Þ. xk , m = 1, 2, 3, 4

(6:32)

μCm′ ðxk Þ. xk

μCm ′ ðxÞ is the value of the membership function for x on Um.

6.2.5 Simulation Experiments 1.

Hungarian algorithm and Takagi-Sugeno fuzzy-logic-based Interbay scheduling method This section presents a simulation experiment to evaluate the effectiveness of the proposed AMPHI dispatching strategy and provides a comprehensive comparison to traditional material handling strategies. Discrete event simulation software eM-Plant is used to model the 300 mm SWFS and the dispatching strategies. The system data used for simulation modeling in this study was shared by a semiconductor manufacturer in Shanghai, China. In the system, there are 216 tools belonging to 54 distinct machining tool groups. There are 22 stockers in the

154

6 Scheduling in Interbay automated material handling systems

Interbay material handling system, and each stocker connects with an Intrabay system. The distance between adjacent stockers is 20 m and the Interbay rail loop is 440 m long with four shortcuts and eight turntables. Each automated vehicle can load only one wafer lot at a time. The speed of the automated vehicle is 1.0 m/ s and the average loading/unloading operation time is 5 s. Three types of wafer lot (jobs A, B and C) are being processed in the system. Totally, six scenarios were designed for different combinations of loading ratios (90% or 100% of the specified system capacity) and automated vehicle numbers (8, 10 or 12 vehicles). Each scenario was replicated in three simulation runs, and each run simulated a production period of 120 days with a transient period of 20 days. The lead times of job A, B and C in each scenario were set as 30, 28 and 31 days, respectively. Simulation experiments were carried out to comprehensively compare the effectiveness of the proposed AMPHI approach and five traditional singleattribute and multi-attribute heuristic dispatching policies. These traditional dispatching approaches including Hungarian-algorithm-based overhead transporter reassignment (HABOR), cassette-look-ahead bid (CLAB), modified first come first served (M-FCFS), short transport distance (STD) and longest waiting time first (LWT) have been adopted in vehicle dispatching of the Interbay system in SWFSs and been approved to be feasible and effective methods. The HABOR dispatching approach takes the Hungarian algorithm to assign empty vehicle to wafer lots, in which wafer lot’s waiting time and position information are considered. In the CLAB approach, wafer lot with minimal total estimated processing time in several succeeding operations wins the vehicle being idle. In M-FCFS, the available vehicles are dispatched to the wafer lots that have the earliest request. If there is another wafer lot (called new call) waiting to be transported at the same stocker when one wafer lot (called old call) at a certain stocker is assigned to a vehicle, then the corresponding request time associated with the new call is set equal to the time when the old call was assigned. The STD dispatching rule means that the wafer lot whose position is the closest to the designated empty vehicle is assigned to the vehicle. In the LWT-based dispatching rule, the wafer lots with the longest waiting time in stockers are assigned to empty vehicles first. A total of eight performance indexes – wafer lot’s cycle time, system throughput, wafer lot’s due date satisfaction rate, system WIP, wafer lot’s movement, delivery time, transportation time and vehicle’s utilization rate – of the AMHS and the SWFS were compared. In the simulation study, AMPHI ranked first among the six dispatching approaches in terms of cycle time, throughput, due date satisfaction rate, WIP level, job movement and transportation time in all six scenarios. AMPHI ranked first in five out of six scenarios in terms of delivery time, and ranked first in four out of six scenarios in terms of vehicle utilization. For clarity and brevity, only the results with 100% loading ratios are listed in Tables 6.5 and 6.6. To further verify the effectiveness of AMPHI approach, the desirability function D(X) is computed. Using the desirability function, the

155

6.2 AMPHI-Based Interbay Automated Material Handling Scheduling Methods

Table 6.5: Performance comparison of different dispatching policies (part 1). Scenario

Dispatching Approach

Throughput Cycle time (s) (lot)

Movement (lot)

Delivery time (s)

Loading ratio = 100%, vehicles = 8

AMPHI HABOR CLAB M-FCFS STD LWT

A (2523) B (2215) C (2147) B (2215) D (2026) C (2144)

A (1794486) B (2206627) C (2527311) B (2220757) D (2631832) C (2511691)

A (589712) A (586410) D (413480) A (586410) B (563111) C (476486)

A (521.7) A (521.2) C (623.2) D (770.6) BC (601.3) B (579.5)

Loading ratio = 100%, vehicles = 10

AMPHI HABOR CLAB M-FCFS STD LWT

A (2581) B (2218) C (2152) B (2221) D (2016) B (2218)

A (1700622) C (2200471) B (2124585) C (2212919) D (2612980) C (2205469)

A (585098) A (586403) C (503901) A (586511) B (560315) A (586468)

B (415.59) A (402) C (450.5) BC (426.28) D (499.47) E (611.68)

Loading ratio = 100%, vehicles = 12

AMPHI HABOR CLAB M-FCFS STD LWT

A (2543) B (2415) C (2358) D (2220) F (2028) E (2142)

A (1733301) B (1831976) C (1989665) D (2206790) F (2574916) E (2318331)

A (597908) A (596433) B (584596) B (586490) C (563404) B (586538)

A (328.93) A (321.22) C (476.85) B (346.98) D (492.18) B (348.79)

Table 6.6: Performance comparison of different dispatching policies (part 2). Scenario

Dispatching Transport approach time (s)

Due date WIP in satisfaction (%) Interbay (lot)

Vehicle utilization (%)

Loading ratio = 100%, AMPHI vehicles = 8 HABOR CLAB M-FCFS STD LWT

A (191.5) A (192.4) A (196.4) A (191.3) B (209.7) A (191.8)

A (0.9841) B (0.6433) C (0.6040) B (0.6379) D (0.5237) B (0.6378)

A (874) B (1179) C (1248) B (1182) D (1367) C (1255)

A (0.768) A (0.7614) E (0.5342) D (0.5688) B (0.728) C (0.6167)

Loading ratio = 100%, AMPHI vehicles = 10 HABOR CLAB M-FCFS STD LWT

A (196.63) A (196.27) B (200.87) A (193.54) A (195.91) A (197.16)

A (0.9895) C (0.6456) B (0.6780) C (0.6380) D (0.5278) C (0.6411)

A (812) B (1177) B (1248) B (1176) C (1384) B (1175)

B (0.6207) A (0.6456) D (0.5325) A (0.638) C (0.5957) B (0.6237)

Loading ratio = 100%, AMPHI vehicles = 12 HABOR CLAB M-FCFS STD LWT

A (200.23) A (199.86) B (205.54) A (196.57) A (201.49) A (200.79)

A (1.0000) B (0.7521) C (0.7307) D (0.6387) F (0.5350) E (0.5901)

A (851) B (986) C (1035) D (1179) E (1369) DE (1251)

A (0.5247) A (0.5236) A (0.5268) B (0.5183) B (0.51) B (0.5191)

156

6 Scheduling in Interbay automated material handling systems

Table 6.7: ANOVA of the AMPHI and traditional HABOR, CLAB, M-FCFS, STD, and LWT dispatching approach. No.

Proposed approach

Traditional approaches

Sum of squares

Degrees of freedom

Mean squares

F-value

Significant differences level

1 2 3 4 5

AMPHI

HABOR CLAB M-FCFS STD LWT

0.48908 1.22688 0.7646 2.13051 0.92152

35 35 35 35 35

0.09314 1.0621 0.38126 2.06919 0.706

8.0 219.15 33.82 1147.16 111.37

p < 0.01 p < 0.01 p < 0.01 p < 0.01 p < 0.01

1

Desirability

0.8

AMPHI HABOR CLAB M-FCFS STD LWT

0.6

0.4

0.2

0 L100V8

L100V10

L100V12 L90V8 Scenario

L90V10

L90V12

Figure 6.8: Comparison of the desirability values of the AMPHI, HABOR, CLAB, M-FCFS, STD and LWT dispatching approaches.

desirability values of the AMPHI, HABOR, CLAB, M-FCFS, STD and LWT dispatching approaches in different scenarios are calculated. Figure 6.8 illustrates that the proposed AMPHI approach clearly outperforms the other traditional dispatching approaches in terms of comprehensive performance in all six scenarios. A statistical analysis of variance (ANOVA) was tested on the desirability values. Table 6.7 shows that AMPHI is significantly different (p ≤ 0.01) from other HABOR, CLAB, M-FCFS, STD and LWT dispatching approaches. The reasons for suggesting the proposed dispatching approach for complex semiconductor wafer fabrication are briefly summarized as follows. (1) The

6.2 AMPHI-Based Interbay Automated Material Handling Scheduling Methods

157

Vehicle’s transportation distance lot’s due date, lot’s waiting time and lot’s origin-destination buffer status are thoroughly considered and used as the key parameters in the wafer handling cost function model. The weight coefficients of these parameters are adaptively adjusted according to the dynamic material handling environment. Therefore, wafer lots with worse cost value can be given priority to transport first in real time. (2) The proposed fuzzy-based Hungarian algorithm is proven as a rapid and effective approach to obtain the optimal or near-optimal vehicle dispatching solutions. The vehicle blockage phenomenon in Interbay material handling can be largely reduced. (3) The proposed fuzzy-based Hungarian algorithm can respond to events such as arrival and transport of the wafers in real time in order to dynamically respond to the dynamic environment of the Interbay material handling system. 2.

Modified Hungarian algorithm and Mamdani fuzzy-logic-based Interbay scheduling method To evaluate the MHA and Mamdani fuzzy-logic-based Interbay scheduling method, discrete event simulation models are constructed using the eM-Plant software. The layout of the 300 mm wafer fab is shown in Figure 6.9. There are 14 bays and 127 sets of equipment for 23 types of process (Table 6.8). The layout parameters of the Interbay material handling system are summarized as follows: the Interbay rail loop is 370 m; the length of the vertical guide is 5.5 m; the vehicle speed is 1 m/s; the time on turntable is 10 s and the loading/ unloading time is 5 s. Three types of wafer lot (jobs A, B and C) are being processed in the system, and the quantity is equal to each other. Figure 6.10 shows the eM-Plant simulation models.

Figure 6.9: The layout of the 300 mm Wafer Fab.

158

6 Scheduling in Interbay automated material handling systems

Table 6.8: The fabrication data and key equipment of bays. Stocker

Bay

Process

Number of critical equipment

Processing time

Capacity (lot/ cassette)

Stoker-1

Bay-1

Cleaning WIS Photo-etching Sputtering 1

3 4 3

20 min 73 min 35 min

2 3 2

Stoker-2

Bay-2

Short circuit test Sputtering 2 Sputtering 3 Chemical vapor deposition

3 1 1 5

44 min 28 min 26 min 43 min

1 2 2 2

Stoker-3

Bay-3

Cleaning WCW Chemical vapor deposition

3 5

19 min 41 min

2 3

Stoker-4

Bay-4

Auto-appearance Cleaning WCW

2 4

48 min 21 min

1 3

Stoker-5

Bay-5

Auto-appearance Sputtering 2 Sputtering 3

1 1 2

48 min 28 min 26 min

1 2 2

Stoker-6

Bay-6

Photo-etching Stripping

5 4

34 min 24 min

2 2

Stoker-7

Bay-7

Auto-appearance Repair (resection) Cleaning WCW Repair (connection)

3 5 1 2

114 min 40 min 22 min 10 min

2 1 2 1

Stoker-8

Bay-8

Dry etch DEC Dry etch DEI Dry etch DCH Auto-appearance Array inspect

1 1 1 2 2

46 min 41 min 53 min 48 min 91 min

2 2 2 1 2

Stoker-9

Bay-9

Auto-appearance Repair (connection) Appearance inspect

3 3 5

48 min 10 min 29 min

1 2 1

Stoker-10

Bay-10

Auto-appearance Repair (resection) Cleaning WCW Wet etch WEI Annealing

3 3 1 1 2

114 min 34 min 20 min 41 min 101 min

2 1 2 3 5

Stoker-11

Bay-11

Cleaning WCW Wet etch WEI

1 2

18 min 41 min

2 3

159

6.2 AMPHI-Based Interbay Automated Material Handling Scheduling Methods

Table 6.8: (continued ) Stocker

Bay

Process

Number of critical equipment

Processing time

Capacity (lot/ cassette)

Stoker-12

Bay-12

Wet etch WEG Stripping Wet etch WED Dry etch DEI

1 3 1 1

28 min 24 min 34 min 41 min

2 2 2 2

Stoker-13

Bay-13

Wet etch WEG Wet etch WED Dry etch DEI Dry etch DCH Dry etch DEC

2 4 1 2 2

28 min 34 min 41 min 53 min 46 min

2 2 2 3 2

Stoker-14

Bay-14

Cleaning WCW Dry etch DEI Photo-etch Dry etch DEC

11 1 7 5

20 min 41 min 79 min 46 min

1 2 4 2

Figure 6.10: The simulation model built by eM-Plant.

160

6 Scheduling in Interbay automated material handling systems

The MHA and fuzzy logic control algorithm are realized by using VC++ and embedded into the simulation models as a dynamic link library (DLL). Simulation experiments were carried out to comprehensively compare the effectiveness of the proposed dispatching method and five traditional single-attribute and multi-attribute heuristic dispatching policies (Table 6.9). These traditional EDD, CLAB, FEFS, STD and RLWT dispatching rules and strategies have been adopted in vehicle dispatching of the Interbay system in SWFSs. The measurements of performance include delivery time, cassette waiting time, system throughput, total delivery amount, mean process cycle and vehicle utilization. Totally, four scenarios were designed for different combinations of two loading ratios (3 cassettes/2.5 h and 3 cassettes/2.75 h) and automated vehicles numbers (8 and 12). Each scenario was replicated in three simulation runs, and each run simulated a production period of 150 days with an aging time Tmax = 2,000(s). The simulation results are shown in Tables 6.9–6.12. In scenarios 1 and 2, the proposed MHAFLC has better performance in terms of mean delivery time, mean waiting time, system throughput and total delivery amount. As for the other two indexes, mean process cycle and vehicle utilization, those dispatching approaches have no significant difference. In scenarios 3 and 4, the proposed MHAFLC has better performance in terms of all measurements except the vehicle utilization. The conclusion can be obtained that when the system load is at a high level, the proposed MHAFLC approach may significantly improve the system performance. Compared with scenarios 3 and 4, the mean delivery times for all dispatching approaches in scenarios 1 and 2 are longer, which indicate that when the system is at a high loading level, the temporary blockages become more frequent. For further analysis, an indicator of the comprehensive function D is introduced to realize the comparison of various dispatching rules. The comprehensive function D is defined as ( Dp ðxÞ =

 * p X   )1=p fu − fu ðxÞ fv ðxÞ − fv− p wu * + wv fu − fu− fv* − fv− u2U v2V

X

(6:33)

Here, U and V denote the sets of performance measures corresponding to the maximum and minimum objectives, and wu and wv are weight coefficients for different measurements. fu* = maxx2X ðfu ðxÞÞ and fu− = minx2X ðfu ðxÞÞ are the permitted maximal and minimal values for index u. fv* and fv− are defined in the same way. Figure 6.11 shows the D-function values of the proposed dispatching method and other rules p = 1. Table 6.12 reports the ANOVA results of different dispatching methods. Furthermore, it is found that in all scenarios the proposed MHAFLC dispatching method has better performance except the vehicle utilization. In terms of D-function value, MHA-based dispatching methods have the best performance in comparison with the other rules. However, the fluctuation of its D-function

6.2 AMPHI-Based Interbay Automated Material Handling Scheduling Methods

161

Table 6.9: Performance of different dispatching methods under scenario S1 (3 C/3 h, OHT = 12). Dispatching strategy

MHAFLC MHA EDD CLAB FEFS STD RLWT

Delivery time (s)

Mean

Deviation

328.9 321.2 333.6 356.8 346.9 492.1 348.7

45,207 45,273 46,303 48,320 49,320 49,981 45,766

Waiting time (s) Throughput (cassette)

Mean cycle time (s)

Delivery amount (cassette)

Vehicle utilization

1,804,917 2,053,857 2,388,146 2,319,167 2,357,323 2,475,098 2,522,490

586,110 587,250 586,782 586,909 586,909 564,088 579,538

0.5247 0.5236 0.5114 0.5267 0.5182 0.5100 0.5190

Mean Deviation 128.7 121.4 133.8 141.3 150.4 290.8 148.0

5,220 5,109 6,455 7,903 8,110 6,782 8,756

2,543 2,415 2,191 2,158 2,220 2,028 2,142

Table 6.10: Performance of different dispatching methods under scenario S2 (3 C/2.5 h, OHT = 12). Dispatching strategy

MHAFLC MHA EDD CLAB FEFS STD RLWT

Delivery time (s)

Mean

Deviation

328.1 325.4 332.6 362.1 351.5 494.0 348.8

45,414 45,615 46,334 48,271 49,330 49,181 45,966

Waiting time (s) Throughput (cassette)

Mean cycle time (s)

Delivery amount (cassette)

Vehicle utilization

978,251 984,410 113,694 1,063,915 1,000,430 1,649,776 1,133,740

589,625 589,584 577,329 582,040 589,341 564,088 579,538

0.6251 0.6366 0.5084 0.5245 0.5214 0.5120 0.5243

Mean Deviation 128.5 125.1 132.2 156.5 196.8 292.0 147.5

5,128 5,201 6,725 7,741 7,910 6,397 8,636

2,593 2,544 2,489 2,528 2,589 2,369 2,505

Table 6.11: Performance of different dispatching methods under scenario S3 (3 C/3 h, OHT = 8). Dispatching Delivery time (s) strategy

MHAFLC MHA EDD CLAB FEFS STD RLWT

Mean

Deviation

521.7 521.2 654.3 641.2 770.5 601.2 659.5

45,414 45,615 46,334 48,271 49,330 49,181 45,966

Waiting time (s) Throughput (cassette)

Mean cycle time (s)

Delivery amount (cassette)

Vehicle utilization

998,745 1,006,923 2,366,741 3,061,225 1,059,187 1,621,700 2,658,287

589,746 589,036 475,757 414,020 587,761 557,276 475,631

0.7258 0.7657 0.6019 0.5319 0.7596 0.7215 0.6167

Mean Deviation 330.2 328.7 353.9 425.8 579.2 491.5 379.2

5,128 5,201 6,725 7,741 7,910 6,397 8,636

2,589 2,587 1,751 1,261 2,577 2,332 1,673

162

6 Scheduling in Interbay automated material handling systems

Table 6.12: ANOVA results of different dispatching methods. Proposed approach

Traditional approaches

1 2 3 4

MHA RLWT STD CLAB

Square deviation (SD) × 1,000

Freedom

Mean SD × 1,000

F-value

(221.5, 1112.5) (189.1, 914.1) (242.7, 1201.3) (88.9, 612.1)

(1, 23) (1, 23) (1, 23) (1, 23)

(221.5, 48.4) (189.1, 39.7) (242.7, 52.2) (88.9, 26.6)

4.58 > F0.05 ð1, 23Þ 4.76 > F0.05 ð1, 23Þ 4.65 > F0.05 ð1, 23Þ 3.34 > F0.05 ð1, 23Þ

Comprehensive function D

0.9 0.8 MHAFLC

0.7 0.6

MHA

0.5

FEFS

0.4

STD

0.3

CLAB

0.2

RLWT

0.1 0 Scenario-1

Scenario-2

Scenario-3

Scenario-4

Figure 6.11: Comparison of D-function values of proposed dispatching method and other rules.

value indicates that the robustness needs to be strengthened. The proposed fuzzy-logic-based control is applied to dynamically adjust the weight and therefore improve the robustness.

6.3 Composite Rules-Based Interbay System Scheduling Method Over the last few years, automated vehicle dispatching in material handling systems has received considerable attention. However, most of traditional vehicles dispatching approaches are usually utilized for single-objective optimization, which only consider the wafers’ process time and waiting time, vehicles’ location and travel time, and so on. In order to satisfy the demand of multiple-objective optimization, some other factors which radically influence the overall system performance and customer satisfaction, such as factors regarding wafer cassettes’ priority, have to be considered. On the other hand, the Interbay material handling system is a dynamic and stochastic system in real-life semiconductor fabrication lines. Static attribute weights in traditional dispatching rules limit the ability for the material handling systems to manage unexpected environmental changes. Therefore, how to deal with

163

6.3 Composite Rules-Based Interbay System Scheduling Method

the stochastic events such as the unexpected vehicle blockages needs to be answered. In this section, a multiple-objective scheduling model of Interbay AMHS is established [2, 3]. Aiming to meet the demands of dynamic adjusting and multipleobjective optimization, a genetic programming (GP)-based algorithm is proposed to generate composite dispatching rules (CDR).

6.3.1 Global Optimization Model of the Interbay System Scheduling Problem The purpose of the scheduling for the Interbay material handling system is to allocate different vehicles to wafer cassettes so that the production and transportation efficiency can be optimized. Such a scheduling problem is actually a special dispatching problem where at any scheduling time t, there are n waiting wafer cassettes and m vehicles. In this section, the scheduling of Interbay material handling is accomplished in two steps. The first step is to prioritize the waiting cassettes by using CDR. The second step is to dispatch the vehicles with the status of “Idle” or “Retrieval” to cassettes according to specially designed transportation strategies and policies. The whole process of CDR-based scheduling method is shown in Figure 6.12. Then, the mathematical model for the scheduling process in Interbay material handling systems is formulated. Index and set M set of waiting cassettes at time t, |M| = m; K set of vehicles at time t, |K|= n; Y set of rails and turntables; S set of stockers; s set of loading/unloading events, PDs = fðk, iÞj∀k, i 2 Mg; PD h fixed loading/unloading time; Vehicles (m)

Wafer cassettes (n)

d1j1 = argmin d1j j∈K

d1j1

d2j2 = argmin d2j

...

j∈K j≠j1

j1

... ...

n >=m

...

Priority

i=1

Wafer cassettes (n)

Vehicles (m)

i=1

Priority

1.

... ... ...

n tRi Þ, ∀j 2 K

(6:40)

i

tDi Xkjp + CPi ≤ ti P Xt D ij , ∀ðk, iÞ  PDs , LðiÞ = LðkÞ ^ WTSðiÞ = WTSðkÞ, ∀j 2 K ðiÞ

(6:41)

tRL + 1 ≤ tPi , ∀i 2 M

(6:42)

E1t < E2t < ... < ENt , ∀t 2 ½0, T 

(6:43)

M X

etyj ≤ 1, ∀t 2 ½0, T , ∀y 2 Y

(6:44)

j=1

Among these constraints, formulation (6.36) ensures that a delivery request can be responded by only one vehicle; formulation (6.37) means that a vehicle can handle only one delivery task at one time; formulation (6.38) guarantees that the vehicle is not available until the current delivery task has been finished; formulation (6.39) indicates the constraint between the loading and unloading times of a cassette; formulation (6.40) indicates the constraint between the loading and unloading times of different cassettes in the same stocker; formulation (6.41) indicates the constraint between the unloading and loading times of a cassette in the same stocker; formulation (6.42) ensures that the loading of the cassette is later than the arrival of the cassette; formulation (6.43) describes the location constraints of vehicles and formulation (6.44) guarantees that there is only one vehicle on a rail at any time.

166

6 Scheduling in Interbay automated material handling systems

6.3.2 Architecture of Composite Rules-Based Interbay System Scheduling Method The compound heuristic rule refers to a compound scheduling rule constructed by taking various heuristic information into consideration, such as various factors influencing the objective function and integrating human visual experience. It is possible to obtain the approximate optimal solution while satisfying the constraints of the optimization problems. To construct composite heuristic rules, two requirements should be satisfied: (1) be able to combine advantages of some single-heuristic rules, and does not require specific parameters related to the scene, with strong robustness and adaptability; (2) dynamic adjustment characteristics. It is able to make the appropriate adjustments according to changes of system optimization objectives at each stage. Figure 6.13 shows the composite heuristic rules-based scheduling framework for an Interbay material handling system. Single-heuristic rules are combined together to obtain a variety of possible composite heuristic rules. With these composite heuristic rules, the scheduling process of an Interbay material handling system is carried out to obtain major transport performances and processing performances, which include wafer lot’s average delivery time, due date satisfactory rate, average cycle time and wafer throughput. These indicators are used as the evaluation indexes of composite heuristic rules in order to analyze its effectiveness at different stages.

R1 Single-heuristic rules R = f (IPT, RPT, LRT, DUE, DIS, AT, t)

Composite heuristic rules generation algorithm Wafer lot’s average delivery time Due date satisfactory rate

R2 R3 . . . Rn

HRm = Rj + Rj+...... Average cycle time Wafer throughput

HRm Interbay material handling system

Figure 6.13: A composite heuristic rule-based scheduling framework of an Interbay material handling system.

6.3 Composite Rules-Based Interbay System Scheduling Method

167

6.3.3 Genetic Programming-Based Composite Dispatching Rule Algorithm In this section, genetic programming-based composite rule generator (GPCRG) is introduced. In GPCRG, several system performance measures are grouped and the solutions with best fitness are obtained by searching from the tree-structure space of a feasible solution. These solutions correspond to different CDR. According to these rules, waiting cassettes are prioritized and vehicles are then assigned. The flowchart of the proposed GPCRG is depicted in Figure 6.14. The procedure includes encoding, initialization, fitness selection, cross and mutation, and stopping criterion. Details of the algorithm are presented in the following sub-sections. 1. Encoding scheme First of all, the set of terminals is defined in Table 6.13 and four fundamental operators are defined as Operator Set = { +, –, ×, \}. In order to obtain the CDR, a binary tree encoding scheme is used in the proposed GPCRG. The above terminals and operators are set as the nodes. It would be noticed that there are no child-nodes for all the terminals while all the operators are parent-nodes.

Set basic parameters Start Encoding Define terminal & operator set according to scheduling objectives Define encoding tree structure

Generate initiate solutions randomly, which include simple dispatching rules

End

Output k best solutions

Stochastic tournament operator

Yes Crossover and mutation operators No

Stop?

D-best local search

Fitness evaluation

Figure 6.14: The flowchart of GPCRG.

168

6 Scheduling in Interbay automated material handling systems

Table 6.13: Terminal set of GPCRG. Variable

Explanation

IPT RPT LRT DUE DIS AT T

Start time of the next operation for the wafer cassette Sum of processing time of all unfinished operations Release time of the cassette Due date of the wafer cassette Distance between the cassette and the stocker Arrival time of cassette to stocker Current time

+

AT

/



IPT

t

LRT



LRT

+

DUE

×

\

t

DUE

\

IPT

t

Figure 6.15: The diagram of coded binary tree of dispatching rules.

The diagrams in Figure 6.15 indicate the rules: Rule -1 = (DUE – t)/[(t + IPT) *LRT] and Rule -2 = [(IPT – t)*LRT]/DUE + AT. 2.

Evolution process (1) Population initialization The initial population of individuals is generated randomly. In order to keep the advantages of the simple dispatching rules, some basic rules such as MFEFS, STD, LWT and EDD are added to the initial population. The CDR with higher performance can be obtained as the coded binary tree evolves. (2)

Genetic operators A. Stochastic tournament operator The procedure is as follows: choose two individuals rando mly and compare their fitness. The better one will be kept into next generation. Repeat this step for M times (suppose the population size is M); a new generation is then developed. This method will balance the computational efficiency and the keeping of good genes.

169

6.3 Composite Rules-Based Interbay System Scheduling Method

B. Crossover operator In general, the crossover operator is regarded as a main genetic operator and the performance of the genetic algorithms depends, to a great extent, on the performance of the crossover operator used. Conceptually, the crossover operates on two chromosomes at a time and generates offspring by combining features of both chromosomes. In the proposed GP-based method, the procedure of crossover operator is shown in Figure 6.16. C. Mutation operator The aim of mutation is to introduce variability and diversity to the population so that the algorithm is able to escape from a local optimum. Two types of mutation exist in the proposed GA-based method: standard mutation and swap mutation. The standard mutation is to replace a child-tree randomly. The swap mutation is to exchange two child-trees in the individual. The procedure of the mutation is shown in Figure 6.17. (3)

3.

Stopping criterion Stopping criterion is a GP parameter which is used to control the termination of the genetic iterative process. Normally, a given number of generations are set as the stopping criterion.

Local search strategy Traditional GP algorithms are easy to fall into the local optimal which results in premature convergence. To prevent this case, a local search strategy is introduced into the GPCRG. The efficiency of the local search algorithms depends on the generating mechanisms and search strategy. In this section, the λ-Interchange local search method is adopted.

+

/

AT

+

DUE

+

Crossover LRT

×

+

LRT

×

Ip /

DUE



DUE

AT t

IPT Ip

t Iq

DUE



\ \ IPT

LRT t

LRT

Figure 6.16: The crossover operator of GP-based CDR generating algorithm.

t Iq

170

6 Scheduling in Interbay automated material handling systems

+ AT

\

LRT t

IPT

PT LRT

× –

/

Standard mutation

+

t

Ip

Swap mutation

AT LRT

+ IPT

+

+ \ LRT

– IPT

t

RPT

DUE

Ip

AT

\ LRT

RPT



DUE IPT

t

Ip

Ip

Figure 6.17: The mutation operator of GP-based CDR generating algorithm.

The λ-Interchange local search improves the quality of the solution by swapping child-trees from different individuals. In the proposed GP-based CDR generating algorithm, only λ = 1, λ = 2 and λ = 3 are considered, which means that between different individuals at most three child-trees can be swapped. According to the value of λ, three child-tree swap operators are defined: (1, 1), (2, 2), (3, 3). For example, the operator (2, 2) on an individual pair (Ip, Iq) indicates that choose two child-trees from individual Ip and add them into individual Iq; meanwhile choose two child-trees from individual Iq and add them into individual Ip. The procedure is shown in Figure 6.18. During the procedure, only when the fitness of the individual is improved this swap is accepted. Since there are normally more than one child-tree in one individual, two strategies are applied to help choose the swap child-trees. (1) First-Search (FS) strategy: choose the first child-tree which improves the fitness by searching from the neighborhood of current solution Nλ ðSÞ. (2) Global-Best-Search (GS) strategy: choose the child-tree which has the best performance for improving the fitness by searching from the neighborhood of current solution Nλ ðSÞ. Although GS strategy performs better in terms of obtaining the global optimal, the computational efficiency decreases. In this section, a compromised d-best strategy is proposed. d child-trees able to improve the fitness are searched from the neighborhood of the current solution and the best one will be chosen as the swap child-tree. 4.

Fitness evaluation Each individual can be decoded into corresponding CDR. In order to measure the performance of the individual, a method based on from-to-table is used. Suppose the CDR decoded from an individual x are named Rule-x and there are totally S stockers in the Interbay material handling system. Before the fitness evaluation, some assumptions are made as follows:

171

6.3 Composite Rules-Based Interbay System Scheduling Method

+

+

/

AT

\

×

+

DUE

AT

+

– PT

LRT

+

/

d-Best-based λ-Interchange

\

+

+



+

IPT

t

×

LRT

Local search method t

IPT

Ip

RPT

DUE

DUE

t

IPT Iq

PT

IPT

Ip

t

DUE

RPT Iq

Figure 6.18: The local search operator (2, 2).

(1)

In each time period T, the total number of requests from wafer cassettes is fixed. The arrival time of the requests is a Poisson process, which indicates that the time interval is exponentially distributed. (2) Each vehicle can only deliver one wafer cassette at a time. The loading and unloading times obey the standard normal distribution. (3) Delivery tasks are evenly assigned to vehicles. (4) The movements of different vehicles are independent of each other. (5) The possibility that the vehicle j is blocked by vehicle j ′ on the downstream transport rail at time t is proportional to the total quantity of vehicles and the material flow rate between stockers on basis of the from-to-table. It can be formulated as ρj = a  M + b  FRðrj , rj′ Þ , where M indicates the total quantity of vehicles, FRðrj , rj′ Þ is the material flow rate between stockers, and a and b are the weights. (6) There exist stochastic blockages of vehicles in the system and the block time obeys the Weibull distribution. This is because if the rail is blocked for a longer time, it is more possible that the rail becomes idle in the follow-up time. For any rail l, αlt < αlt′ , t < t′, αlt = 1 − expð − ððt − t0 Þ=λÞk Þ indicates the possibility that the rail l is blocked before time t and becomes idle after time t, tc is the time when the rail l becomes blocked, and k and λ are shape parameters of the Weibull function. Therefore, the expectation of the block time of rail l at time tl when the vehicle enters rail l can be obtained by the following formulation: BlockTimelt =

∞ X

ið1 − αlt Þi − 1  αlt = 1=αlt

(6:45)

i=1

The evaluation of individual fitness takes into account three performance measures: average delivery time, mean processing cycle time and waiting time of the cassette. The evaluation procedure consists of following steps: Step 1. Calculate the arrival time tRi for each stocker according to the probability distribution, and sort all tasks based on their arrival times tR1 ≤ ... ≤ tRm .

172

6 Scheduling in Interbay automated material handling systems

Step 2. Calculate the starting times of all delivery tasks according to tRi . Suppose the set of arrival times of new tasks is TR = ½tR1 , tR2 , ..., tRm , and the set of times when the vehicles finished the previous tasks is TD = ½tD1 , tD2 , ..., tDm − 1 . At time t 2 TR ∪ TD the dispatching is triggered, calculate the priorities of cassettes according to Rule-x. Step 3. Calculate the delivery time for each vehicle. The delivery time is the sum of fixed moving time and block time. It is formulated as follows: DeliveryTimeij = disðSik , Sip Þ=v +

X

ρj BlockTimel tl

(6:46)

l2Rj

Here, disðSik , Sip Þis the fixed distance between the location k of the current stocker and location p of the target stocker. v is the moving speed of vehicles. BlockTimelt is the j block time of vehicle j on downstream rail l. In a real-life case, the rail constraints ensure that at any time there is only one vehicle on a rail. Therefore, the delivery time has to take into account the temporary block time due to the competence of vehicles. In order to obtain the average delivery time at any scheduling time test, a heuristic-based algorithm is proposed. Suppose there are m = 3 vehicles and n = 4 rails. The sequence of railways for all vehicles in current schedule is recorded as l1l2l3l4 |l2l3l4 |l3l4, and a matrix O[m][n] is defined as 2

3 ð2Þ ð3Þ ð4Þ ð5Þ ð6Þ ð7Þ 5 ð12Þ ð9Þ ð10Þ

ð1Þ O = 4 ð8Þ ð11Þ

(6:47)

In this matrix O, each row represents a sequence of railways through which a vehicle will move in current schedule. Each railway is numbered in accordance with the sequence constraint. The cost of delivery time on railways for each vehicle is addressed in Table 6.14. Sort all elements in the matrix O according to the time cost while the sequence constraints are still kept, and a vector R = {5, 1, 2, 9, 10, 3, 4, 6, 11, 12, 7, 8} can be obtained which indicates the moving sequence of vehicles under the railway constraints. Define matrix TM[n][nm] as follows: each row of TM represents a railway, and each column is corresponding to an element in matrix O. The value of the

Table 6.14: The cost of deliverytime on railways for each OHT. Vehicle 1 2 3

l1

l2

l3

l4

(1)3 (8)0 (11)0

(2)3 (5)2 (12)0

(3)8 (6)5 (9)4

(4)3 (7)3 (10)5

6.3 Composite Rules-Based Interbay System Scheduling Method

173

elements in the matrix TM stands for the delivery time cost. If the element values 0, it means that the vehicle will not move through this railway in current schedule. 2

3 0 0 0 6 0 3 0 0 6 TM = 6 0 2 0 6 0 4 0 0 0 1 ð1Þ ð2Þ ð3Þ ð4Þ

0 0 0 0 0 0 2 0 0 0 2 0 0 5 0 0 0 5 0 0 0 3 0 0 ð5Þ ð6Þ ð7Þ ð8Þ ð9Þ ð10Þ

3 0 0 0 0 7 7 0 0 7 7 0 3 5 ð11Þ ð12Þ

(6:48)

However, in the matrix TM, only the railway constraints can be reflected. The sequence constraints are still unknown from this matrix. In order to obtain the finish time of the delivery task for each vehicle, the matrix TM has to be mapped to another matrix TT[n][nm] as follows: 2

3 0 0 6 0 0 0 6 6 0 TT = 6 0 0 4 0 0 0 ð1Þ ð2Þ ð3Þ

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ð4Þ ð5Þ ð6Þ ð7Þ ð8Þ

3 0 0 0 0 0 0 0 0 7 7 0 0 0 0 7 7 0 0 0 0 5 ð9Þ ð10Þ ð11Þ ð12Þ

(6:49)

In the matrix TT, each row indicates a railway. The element with a non-zero value in column j  n (j = 1, 2, . . ., m) stands for the finish time of the delivery task processed by vehicle j. When generating the matrix TT, precedence constraints have to be considered. If there is no precedence constraint, delivery time = cost time represented by current index R(i) + max (all elements of row l in matrix TM) + stochastic block time. If there exist precedence constraints, delivery time = cost time represented by current index R (i) + max (max (all elements of row l in matrix TM), max (all elements of row a in matrix TM)) + stochastic block time. l is the row index of the non-zero element in column R(i) in matrix TM, a is the row index of the non-zero element in column R(i) – 1 in matrix TM. Based on the above algorithm, the matrix TT can be obtained as follows: 2

3 0 6 0 6 6 TT = 6 0 0 6 4 0 0 ð1Þ ð2Þ

3 0 0 0 0 0 22 0 0 9 0 0 0 2 0 0 0 0 0 0 9 7 7 14 0 0 19 0 0 4 0 0 0 7 7 0 17 0 0 22 0 0 9 0 0 5 ð3Þ ð4Þ ð5Þ ð6Þ ð7Þ ð8Þ ð9Þ ð10Þ ð11Þ ð12Þ

(6:50)

The average delivery time of all tasks at scheduling time toss can be calculated:  = ð17 + 22 + 9Þ=3 = 16. D The pseudo-code of the proposed algorithm for calculating the finish time of all tasks under the railway constraints and stochastic blockages is presented as follows:

174

6 Scheduling in Interbay automated material handling systems

Algorithm 6.1. Algorithm for calculating the finish time of all tasks under the railway constraints and stochastic blockages TT = 0; j = 0; For i = 1 to n //First column in matrix TT TT [i][R(1)] = TM[i][R(1)]; For i = 2 to

m P j=1

Xij = 1//Other columns in matrix TT

{ //row index of the non-zero element in column R(i) in matrix TM l = Find NonZeroIndex(TM, i); //row index of the non-zero element in column R(i) – 1 in matrix TM a = Find NonZeroIndex(TM, R(i) - 1); If (i%n == 1) //no precedence constraint Then {j ++; TT[l][R(i)]] = TM[l][R(i)] + MaxRowElem(TT, l) +pj *BlockTime; //MaxRowElem returns the maximum of all row elements //pj is the possibility that vehicle j meets block } Else //with precedence constraints {TT[a][R(i)]] = TM[a][R(i)] + max(MaxRowElem(TT, l), TT(a, R(i) – 1)) +pj *BlockTime; } }

At any scheduling time ts, the average delivery time for all tasks is formulated as follows: m 1X t DT ðxÞ = TT½a½j  n, j = 1, 2, ..., n (6:51) m i=1 Here, a is the row index of the non-zero element in columnj  n in the matrix TT. Step 4. Prioritize all delivery tasks according to Rule-x, and calculate the average lateness of the next processing operation for current cassette. For a delivery taski 2 M: Q

PDT t =

Si X

PTj =PCSi

(6:52)

j=1

Here, Si is the location of the target stocker of task i. QSi is the length of the queue in the buffer of the stocker. PTj indicates the processing time of the current operation in the buffer. PCSi represents the maximum processing capacity of key equipment. The average lateness of the next operation for all cassettes can be formulated as

175

6.3 Composite Rules-Based Interbay System Scheduling Method

Q

t

PDT ðxÞ =

Si jM j X X

PTj =PCSi

(6:53)

i=1 j=1

Step 5. Calculate the average waiting time for all cassettes at scheduling time t. t

WT ðxÞ =

jM j X

WaitTimei

(6:54)

i=1

In formulation (6.54), j = fjjj 2 K ^ dij = arg min dij′ g indicates the nearest vehi j′2K

cle

to

the

cassette

of

task

i 2 M, WaitTimei = ðWTmax − WTi Þ=WTmax ,

WTi = tPij − tRij , WTmax = const and WaitTimei is the waiting time for task i 2 M. Step 6. Calculate the fitness of Rule-x after the linear normalization and weighting for formulations (3.28), (3.30) and (3.32). Ft ðxÞ = Ft1 ðxÞ + Ft2 ðxÞ + Ft3 ðxÞ

(6:55)

F i ðxÞ 2 ½0, 1, i = 1, 2, 3, it is obtained by t

Ft1 ðxÞ = ðDT ðxÞ − DT min Þ=ðDT max − DT min Þ t

Ft2 ðxÞ = ðPDT ðxÞ − PDT min Þ=ðPDT max − PDT min Þ t

Ft3 ðxÞ = ðWT ðxÞ − WT min Þ=ðWT max − WT min Þ

(6:56)

(6:57) (6:58)

The fitness function for individual x in the whole scheduling period T can be formulated as FT ðxÞ =

T X

Ft ðxÞ

(6:59)

t=0

The individual is better if the value of its fitness FT ðxÞ is smaller, and has a relatively larger probability to be selected and kept to the next generation. The population is therefore evolved. 6.3.4 Simulation Experiments To evaluate the CDR generated by proposed GPCRG, discrete event simulation models are constructed using the eM-Plant software. The GP-based CDR

176

6 Scheduling in Interbay automated material handling systems

generating algorithm is realized by using VC++ and embedded into the simulation models as a DLL. The established simulation models consist of three feed rates: 3 cassettes/2.5 h (high load), 3 cassettes/2.75 h (high load) and 3 cassettes/3 h (low load). Three cases of vehicle quantity are considered: 8, 10 and 12. The total simulation scenarios therefore count 3 × 3 = 9, and for each scenario the experiments repeat three times. The from-to-table of the current Interbay AMHS is obtained and presented in Table 6.15. Parameters of the proposed GP-based CDR generating algorithm are addressed in Table 6.16. After 100 × 200 = 20,000 evaluations, 5 best individuals are obtained. The represented CDR are listed in Table 6.17. In scenario 1 where feed rate is 3 cassettes/2.5 h and vehicle quantity is 12, the scheduling results are shown in Table 6.18.

Table 6.15: The from-to-table of the current Interbay AMHS. Stocker

STK1

STK2

STK3

STK4

STK5

...

STK14

Total

STK1 STK2 STK3 STK4 STK5 STK6 STK7 STK8 STK9 STK10 STK11 STK12 STK13 STK14

– 4.32 3.28 – – 0.97 – 1.13 0.26 – – – – –

– – – – 3.32 – – 2.54 – – 0.09 – – 2.73

2.67 2.54 – 0.71 0.90 – – – – – 2.33 2.09 1.97 –

0.25 0.67 – – 0.34 1.44 – – – – 1.77 0.38 0.41 –

6.85 0.98 0.22 – – – – – 6.12 4.33 – – 0.28 0.25

... ... ... ... ... ... ... ... ... ... ... ... ... ...

0.17 7.09 – – 3.33 2.91 0.44 – – 0.22 0.14 0.09 – –

10.91 13.48 9.02 3.29 12.15 9.91 2.52 12.24 13.98 6.44 7.05 11.77 9.88 6.97

Total

9.96

8.78

12.31

5.26

19.03

...

14.29

127.93

Unit: Cassette per hour

Table 6.16: The parameter set for the GP-based CDR generating algorithm. Parameter Population size Evolved generation Max depth of initial tree Max depth of crossover tree Possibility of crossover

Value 100 200 10 15 0.3

Parameter Possibility of standard mutation Possibility of swap mutation Possibility of λ-interchange Value of d Weight (a, b)

Value 0.05 0.05 0.1 2 (0.08, 0.05)

177

6.3 Composite Rules-Based Interbay System Scheduling Method

Table 6.17: The generated composite dispatching rules. Composite dispatching rules Formulations    GPCRG-R1 DUE.LRT ðRPT + DUE Þ DUE + RPT =DUE − RPT − AT . −k.DIS + AT − LRT t − AT AT ðRPT + t − AT Þðð2 RPT − t Þ + DUE Þ k= LRT GPCRG -R2 RPT .ðDUE − t Þ DUE − ð2LRT − DUE Þ + ððDUE − t Þ=RPT Þ.DIS   GPCRG -R3 AT 2 ðAT + RPT =ðDUE − t ÞÞ DIS + . IPT ððAT 2 + DUE. AT ÞðRPT + DUE ÞÞ   GPCRG -R4 RPT 1− + DIS. ðt − AT Þ LRT + IPT − t DUE + r ðRPT − AT Þ GPCRG -R5 ,e LRT − t. ðRPT + AT Þðt − LRT Þ l 1=LRT + ðt − DISÞ=AT , r = LRT

Table 6.18: Performance of the generated composite dispatching rules. Composite dispatching rules GPCRG-R1 GPCRG-R2 GPCRG-R3 GPCRG-R4 GPCRG-R5

Mean cycle time (h)

Mean delivery time (s)

Due date satisfaction rate (%)

Throughput (c)

507.62 526.09 516.44 529.58 511.99

325.01 340.87 323.14 339.29 324.90

87.91 82.02 91.22 80.18 89.14

2,539 2,477 2,514 2,482 2,511

Table 6.19: ANOVA results of D-value of generated composite dispatching rules. GPCRG-R1 R2 R3 R4 R5

Square deviation (SD) × 1,000

Freedom

Mean SD × 1,000

F-value

Difference interval

(401.1, 4972.5) (200.7, 9964.1) (325.5, 3910.3) (76.4, 2612.9)

(1, 53) (1, 53) (1, 53) (1, 53)

(401.1, 93.8) (200.7, 179.7) (325.5, 73.8) (76.4, 49.3)

4.33 1.07 4.41 1.55

A B A B

The ANOVA results of D-value of the generated CDR are shown in Table 6.19. The benchmark is GPCRG-R1. “A” and “B” indicate the significant difference interval (confidence level less than 95%). Figure 6.19 shows that GPCRG-R1, GPCRG-R3 and GPCRG-R5 have better comprehensive performance compared with the other two rules, and in addition are relatively more stable. The reason is that such three CDR contain more information of the

Expectation of D

178

6 Scheduling in Interbay automated material handling systems

1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

GPCRG-R1 GPCRG-R3 GPCRG-R5 GPCRG-R2 GPCRG-R4

S-1

S-2

S-3

S-4

S-5

S-6

S-7

S-8

S-9

Figure 6.19: The D-value comparison among the proposed five composite rules in nine scenarios.

3,000,000

Mean cycle time

2,500,000 2,000,000

GPCRG-1 MHAFLC

1,500,000

MFEFS EDD

1,000,000

STD 500,000

RLWT

-9

Sc

en

ar

io

io

-8

-7

ar en Sc

en

ar

ar

io Sc

en Sc

io

-6

-5 ar

io ar

en Sc

en

io

-4

-3 io ar Sc

en Sc

ar en Sc

Sc

en

ar

io

io

-2

-1

0

Figure 6.20: Comparison of mean cycle time among all dispatching methods in nine scenarios.

system. In contrast, GPCRG-R2 does not take into account the factor of due date DUE and GPCRG-R4 does not include the system time t. Some other dispatching rules including MHA, EDD, STD, R-LWT and MFEFS are compared with generated CDR. These traditional dispatching rules and strategies have been adopted in vehicle dispatching of the Interbay system in SWFSs and been approved to be feasible and effective methods. The experimental results are depicted in Figures 6.20–6.22. An indicator of the comprehensive function D is introduced to realize the comparison of various dispatching rules. The variables in function D include average delivery time, wafer cassette throughput, mean processing cycle time and due date satisfaction rate.

6.3 Composite Rules-Based Interbay System Scheduling Method

179

Due-date satisfaction rate

1.2 1.0 0.8

GPCRG-1 MHAFLC

0.6

MFEFS EDD

0.4

STD RLWT

0.2

ar io -9

Sc en

ar io -8

ar io -7

Sc en

Sc en

Sc en

ar io -6

ar io -5

Sc en

ar io -4

ar io -3

Sc en

Sc en

ar io -2

Sc en

Sc en

ar io -1

0

Figure 6.21: Comparison of due date satisfaction rate among all dispatching methods in nine scenarios.

800 700

Mean delivery time

600 GPCRG-R1 500

MHAFLC EDD

400

MFEFS STD

300

RLWT 200 100

-9 io en ar

Sc

en a Sc

en a Sc

rio -8

rio -7

-6 io en ar

Sc

rio -5 en a

Sc

rio -4

Sc e

en a Sc

na

rio -3

-2 io en ar

Sc

Sc e

na

rio -1

0

Figure 6.22: Comparison of mean delivery time in nine scenarios.

In terms of mean wafer processing cycle time and due date satisfaction rate, GPCRG-R1 and MHAFLC have best performance than the other methods in all nine scenarios – especially when the system is under a high load and vehicles are relatively not enough. It is also noticed that when the transport capacity becomes the bottleneck, the generated CDR are able to maximize the improvement of transport

180

6 Scheduling in Interbay automated material handling systems

efficiency. As for the mean delivery time, when the system load is at a lower level, the performance of all scheduling methods improves. GPCRG-R1 and MHAFLC still have best performance than the other methods. However, when the system load is at a high level, MHAFLC performs better than GPCRG-R1.

6.4 Conclusion In this chapter, an adaptive multi-parameter and hybrid intelligence (AMPHI)-based Interbay material handling scheduling method has been proposed based on four parameters: delivery times, wafer lots’ waiting times, wafer lots’ due dates and processing characteristics. A real-time material handling scheduling system has been optimized through a Hungarian algorithm and an MHA. The fuzzy logic control methods (including Mamdani fuzzy inference and Takagi-Sugeno fuzzy inference) have been used to adjust scheduling strategies dynamically according to the realtime system status. At the same time, this chapter has also studied the global optimization scheduling method based on composite heuristic rules, has proposed the global optimization model of an Interbay system scheduling problem and has designed a composite rule algorithm based on GP. To verify the effectiveness of these two proposed scheduling methods, the actual production data of a 300 mm semiconductor wafer fab has been used to establish a discrete time simulation model with eM-Plant software. By analyzing the experimental results, the AMPHI-based Interbay material handling scheduling method has taken into account randomness of the material handling environment and has met the requirements of the multi-objective optimization. However, it has showed inadequacy in real-time scheduling in solving a complex mathematical model. Therefore, this method is suitable for solving Interbay scheduling problems with higher demand of multi-objective optimization and lower real-time requirements. And it is better to adopt the composite rules-based Interbay system scheduling method to solve real-time scheduling problems. The composite rules can also guarantee the global performance to a certain extent. It is suitable for Interbay material handling scheduling problems with high real-time demand and global performance optimization.

References [1]

[2] [3]

Qin, W., Zhang, J., Sun, Y.B. Dynamic dispatching for Interbay material handling by using modified Hungarian algorithm and fuzzy-logic-based control. The International Journal of Advanced Manufacturing Technology, 2013, 67(1 –4): 295–309. Wu, L. Research on Intelligent Scheduling Technologies of AMHS in Semiconductor Wafer Fabrication System. Shanghai Jiaotong University, 2011. Sun, Y. Research on Intelligent Scheduling Technology of Interbay System in Semiconductor Wafer Fabrication System [D]. Shanghai Jiaotong University, 2011.

References

[4]

[5]

181

Lin, J.T., Wang, F.K., Yen, P.Y. Simulation analysis of dispatching rules for an automated Interbay material handling system in wafer fab. International Journal of Production Research, 2001, 39(6): 1221–1238. Qin, W., Zhang, J., Sun, Y.B. Multiple-objective scheduling for Interbay AMHS by using geneticprogramming-based composite dispatching rules generator. Computers in Industry, 2013, 64 (6): 694–707.

7 Scheduling in Intrabay automatic material handling systems The Intrabay material handling system is primarily responsible for wafer handling between processing machines and stockers within the processing bays. Since the Intrabay material handling systems are used to move material directly between processing machines, it is necessary for the Intrabay system to consider the status of each processing machine, which is obviously different from the Interbay system. In the production scheduling process in Intrabay material handling systems, it is first necessary to determine each wafer lot’s priority according to their status. Second, the wafer lots with higher priority are assigned to vehicles according to dispatching rules. Timeliness constraints between processing operations should be considered seriously because Intrabay systems are directly connected with the processing areas. At the same time, the optimization objective in the Intrabay systems focuses on the performances of the processing machines, such as the wafer lot’s average waiting time and average delivery time, which are closely related to the important performance indexes of semiconductor wafer fabrication systems (SWFSs), i.e., wafer lot’s throughput and cycle time. In addition, similar to Interbay system scheduling problems, the Intrabay scheduling systems have a strong dynamic randomness, temporary congestions and deadlocks which should be taken into consideration.

7.1 Intrabay Automatic Material Handling Scheduling Problems An Intrabay material handling system typically consists of transportation rails, load port for stockers and tools, and automatic vehicles (as shown in Figure 7.1). The automotive vehicles are guided by the transportation rail, which is typically a monorail system. Along the rail, automotive vehicles can transport wafer lots and interface with the load ports of stockers and tools. Automatic vehicles are adopted to move wafer lots. In this chapter, overhead hoist transports (OHTs) are selected as the automatic vehicle with which to investigate the vehicle dispatching problem in Intrabay material handling systems in a 300 mm SWFS. When the Intrabay material handling system is running, wafers are fed into the stockers at random time and then transported to the processing machines. After all the processes are completed, the wafers are returned to the stockers leaving the Intrabay material handling system. The automatic vehicles are continuously transported in a one-way sequence on the transportation rails or respond to requests from load ports of stockers and processing machines. https://doi.org/10.1515/9783110487473-007

7.1 Intrabay Automatic Material Handling Scheduling Problems

Tool 1

Tool 3

Tool 5

Tool 7

183

Tool n

OHT

OHT Transportation rails Intrabay system

Load/unload port

Stocker

Move direction

Tool 2

Tool 4

Tool 6

Tool 8

Tool n+1

Interbay system

Figure. 7.1: The layout of an OHT-based Intrabay material handling system.

The wafer carriers waiting for the load port are transported to the next processing machine or stocker. Therefore, the Intrabay material handling scheduling problem can be described as the process of allocating the empty automatic vehicles to transport wafers in the load ports of stockers or processing machines. The Intrabay material handling scheduling problem has the following constraints: 1. Congestion constraints: In the Intrabay material handling system, automatic vehicles are continuously transported in a one-way sequence on the transportation rails. They cannot cross each other, and the downstream automatic vehicles parking for loading/unloading easily lead to the temporary blockage of upstream automatic vehicles. 2. Inter-process time constraints: Wafer waiting times between processing operations have timeliness restrictions (wafers in the last processing machine to complete the processing operation are required in the specified time to enter the next processing machine for processing, which is known as timeliness). 3. Deadlock constraints: A deadlock exists in the Intrabay material handling system (one wafer is jammed by another wafer waiting to be handled at the unloading port of its target processing machine, which causes the wafer to be unloaded after being transported by automatic vehicle to the target processing machine). It is similar to the Interbay material handling scheduling problem. In this chapter, a greedy dynamic priority (GDP)-based Intrabay material handling scheduling method [1] and a pull and push strategy-based Intrabay material handling scheduling method [2, 3] are presented.

184

7 Scheduling in Intrabay automatic material handling systems

7.2 GDP-Based Intrabay Material Handling Scheduling Method 7.2.1 Formulation of the Intrabay Material Handling Scheduling Problem In the Intrabay material handling system, the mathematical model is established with the objectives of minimizing normalized wafer lot’s delivery time, wafer lot’s total delivery time, cycle time, WIP quantity, and maximizing transport amount, vehicle utilization, wafer lot’s throughput, wafer lot’s due date satisfactory rates. 1. Notations Tu – Smallest time unit in the material handling process (vehicle speed, loading/unloading time and wafer delivery request arrival time are integer multiples of Tu); T – Cycle time period in the Intrabay system, T 2 fTu , 2Tu , 3Tu , ...g; w – Within T, the quantity of wafers waiting for transportation in the Intrabay system; q – Within T, the quantity of wafer lots entering the Intrabay system; l – Within T, the quantity of wafer lots leaving the Intrabay system after finishing all processing operations; n – Within T, the quantity of wafers delivered by vehicles; m – The quantity of vehicles in the Intrabay system; Y – Set of tracks and shortcuts in the Intrabay system; S – Set of stockers in the Intrabay system; etyj – etyj = 1 denotes that vehicle j is located at track or shortcuts y at time t, y 2 f1, 2, ..., jY jg; αu – Weighted coefficient of System performance indexes, αu 2 ð0, 1, u = 1, ..., 8; PDs – Set of pick-up/drop-off events at stocker s, PDs = f..., ðk, iÞ, ...g, s 2 S; PrðiÞ – Product type of wafer i; DeðiÞ – Position of destination stocker or processing machine of wafer i; ReðiÞ – Position of stocker or processing machine of the waiting wafer i; d – Maximum time in which a vehicle can transport a wafer lot; Tmax t Tmax – Maximum value of wafer lots’ total transportation time; C Tmax – Maximum value of wafer lots’ total cycle times; Lmax – Maximum value of wafer lots’ throughput; i – Maximum interval time between successive processing steps of wafer lot i; TW

TDi – The shortest delivery time for wafer lot i from the current stocker to the destination stocker; Ejt – Location of vehicle j at moment t; Fi – Due date relaxation coefficient of wafer lot i; Tpi – Current processing time of wafer lot i; h – The time that the vehicle spends in loading/unloading a wafer lot;

185

7.2 GDP-Based Intrabay Material Handling Scheduling Method

tri – The moment that wafer lot i is released to the Intrabay system, tri subject to certain probability distribution; twi – The moment that wafer lot i sends out a transportation request; tpi – The moment that wafer lot i is loaded; tdi – The moment that wafer lot i is unloaded; i – The moment that wafer lot i leaves the Intrabay system after finishing all tout processing operations.

2.

Decision variables Xijt – A 0–1 decision variable where Xijt is equal to 1, if wafer lot i is transported by vehicle j at moment t; otherwise, Xijt is equal to 0.

3.

Objective function: 0

n P

ðtdi − tpi Þ

B i=1 Obj = MinB @α1 n × T d

n P

+ α2

i=1

ðtdi − twi Þ

t n × Tmax

max

l P

+ α5

n P

ðtdi − tpi + 2hÞ ðw − nÞ i=1 + α4 + α3 w m×T 1

v ðtout

− trv Þ

ðLmax − lÞ + α6 v = 1 C Lmax l*Tmax

+ α7

(7:1)

l v ðq − pÞ α8 X ðtout − trv ÞC C + C v q l v = 1 F × Tpv A

4. Constraints: m X

Xijt = 1, Xijt 2 f0, 1g, ∀t 2 ½0, T, i = 1, ..., n

(7:2)

Xijt 2 f0, 1g, ∀t 2 ½0, T, j = 1, ..., m

(7:3)

j=1 n X

Xijt = 1,

i=1 tk

ti

tdk Xkjd + h + 1 ≤ tpi Xijp ti

ti

(7:4)

tpi Xijp + TDi + h ≤ tdi Xijd

(7:5)

i tdi − twi + h ≤ TW

(7:6)

tpk Xkjp + h ≤ tdi Xird , ∀ ReðkÞ = DeðiÞ, ðk, iÞ 2 PDs

(7:7)

tk

ti

186

7 Scheduling in Intrabay automatic material handling systems

tk

ti

tdk Xkjd + Tpk ≤ tpi Xirp , ∀ PrðkÞ = PrðiÞ, ðk, iÞ 2 DPs

(7:8)

tri < twi < tpi

(7:9)

t E1t < E2t <    < Em

m X

etyj ≤ 1 , ∀ t 2 ½0, T, ∀ y 2 Y

(7:10)

(7:11)

j=1

where formula (7.1) is the objective function considering normalized wafer lot’s delivery time, wafer lot’s total delivery time, transport amount, vehicle utilization, wafer’s throughput, cycle time, WIP quantity and wafer lot’s due date satisfactory rate; formula (7.2) ensures that each wafer lot can be transported by only one vehicle; formula (7.3) makes sure that a vehicle can only transport a wafer lot at a time; formula (7.4) ensures that the vehicle can start next transportation task unless it has completed current task; formula (7.5) makes sure that the loading and unloading times of the same wafer lot must satisfy certain constraints; formula (7.6) makes sure that a wafer lot needs to meet the timeliness constraints between processes; formula (7.7) ensures that the loading and unloading times of a wafer lot at a processing machine or stocker must satisfy certain constraints; formula (7.9) makes sure that loading times of different wafer lots in the same processing machine must meet chronological constraints; formula (7.10) ensures the location constraints between vehicles; formula (7.11) makes sure that each track, shortcut or turntable can only carry one vehicle at any time.

7.2.2 Architecture of the GDP-Based Intrabay Material Handling Scheduling Method The architecture of the proposed GDP-based Intrabay material handling scheduling method is shown in Figure 7.2. First, with the advantages of rapid calculating and dynamic identifying of fuzzy set theory, the wafer lots’ dynamic priority decision-making model is developed. Second, a simplified Hungarian method is constructed to assign OHT vehicles to the wafer lots with higher priority in order to reduce the temporary blocking of automatic vehicles during wafer lot

7.2 GDP-Based Intrabay Material Handling Scheduling Method

Lot’s due date satisfaction factor Intrabay system load factor Lot’s wait time factor

Fuzzy logic-based Lot’s transport dynamic priority priority decision-making model

187

Greedy OHT vehicle dispatching policy Simplified Hungarian algorithm-based vehicle assigning model

Greedy vehicle transport policy

OHT vehicle dispatching Intrabay material handling system in a semiconductor wafer fabrication fab

Figure 7.2: Architecture of the GDP-based Intrabay material handling scheduling method.

handling. Finally, according to the material handling optimization assignment results, a greedy vehicle transport policy is proposed to use automatic vehicles to transport wafer lots quickly and efficiently in order to avoid the deadlock in the Intrabay material handling process.

7.2.3 Fuzzy Logic-Based Dynamic Priority Decision-Making Model In this section, the membership functions of the input and output variables are defined by using a triangular membership function. The fuzzy inference policy is determined by the Mamdani inference. And the aggregation policy is determined by the Mini-max rule. The maximal membership function approach is taken as the defuzzified policy. The architecture of the proposed fuzzy logic-based [4–6] dynamic priority decision-making model [7] is shown in Figure 7.3. 1. Input variables The input variables of the fuzzy logic-based dynamic priority decision-making model consist of the Intrabay system load factor, the lot’s due date satisfaction factor and the lot’s wait time factor. Detailed descriptions of these input variables are presented as follows. (1) Lot’s due date satisfaction factor (DDSF) The DDSF is used to measure the current lot’s de date satisfaction. The DDSF is calculated by eq. (7.12):

DDSF =

RTi RPi × Fac

(7:12)

188

7 Scheduling in Intrabay automatic material handling systems

Intrabay system load factor x1

X is A1

y is B1

Rule 1

X is A2

y is B2

Rule 2

Fuzzy inference engine

Fuzzifier

Lot’s due date satisfaction factor x2 Fuzzifier Lot’s wait time factor x3 Fuzzifier

......

Aggregator

X is An–1

y is Bn–1

X is An

y is Bn

Defuzzifier

Lot’s transport priority

Rule n–1 Rule n

Figure 7.3: The architecture of the fuzzy logic-based dynamic priority decision-making model.

Fac =

n X ðCycle timeÞi 1 × PTi n i=1

(7:13)

where Fac is the average value of the historical records for each type lot, and Fac is computed as in eq. (7.13); RTi is the remaining time to the due date of lot i; PTi is the overall processing time of lot i; RPi is the remaining processing time of lot i; and n is the number of historical records for the current type of lot. (2) Lot’s waiting time factor (WTF) The WTF is adopted to measure the current wafer lot’s waiting time. WTF is defined as WTF =

CWi AW

(7:14)

where AW is the average waiting time of the current waiting lot, and AW is n P 1 calculated as AW = CWi × . CWi is the current waiting time for lot i, and n n i=1 is the number of historical records for the current type of lot. (3) System load factor (SLF) The SLF is used to measure the current Intrabay system load ratio. A higher SLF value of the Intrabay system indicates a heavy system load. The SLF is calculated by SLF =

NC NV

(7:15)

where NC is the current number of lots waiting for moving in the Intrabay material handling system and NV is the number of OHTs in the Intrabay material handling system.

189

7.2 GDP-Based Intrabay Material Handling Scheduling Method

2.

Membership functions and the rule table The input variables include DDSF, WTF and SLF. The DDSF can be categorized into bad satisfaction (BS), normal satisfaction (NS) and good satisfaction (GS). The membership function of DDSF is shown in Figure 7.4(a). The WTF can be categorized into not long (NL), medium long (ML) and very long (VL). The membership function of the WTF is shown in Figure 7.4(b). The SLF can be categorized into not heavy (NH), medium heavy (MH) and very heavy (VH). The membership function of the WTF is shown in Figure 7.4(c). The lot’s transport priority is the output variable of the fuzzy logic-based dynamic priority decision-making model. The transport priority can be classified as most high priority (MHP), very high priority (VHP), high priority (HP), low priority (LP) and very low priority (VLP), which indicate levels 9, 8, 7, 6 and 5, respectively. The membership function of the output variable is shown in Figure 7.4(d). Table 7.1 lists the membership functions of three input variables and one output variable. The fuzzy rule table is developed as shown in Table 7.2 in accordance with the simulation experiences of OHT dispatching and the definitions of the variables DDSF, WTF and SLF.

M(DDSF) 1.0

M(SLF) BS

NS

GS

0.7 0.75 0.8 0.85 0.9 0.95 1.0 DDSF

M(WTF)

VH

06

(b)

0.8

1.0

1.2

1.4 SLF

M(P) NL

1.0

ML

VL

VLP

1.0

0 (c)

MH

0

0 (a)

NH

1.0

LP HP

VHP

MHP

0 0.6

0.5

1.0

1.5

WTF

(d)

4

5

6

7

8

9

Priority

Figure 7.4: (a) Membership function of the DDSF. (b) Membership function of the SLF. (c) Membership function of the WTF. (d) Membership function of the lot’s transport priority.

190

7 Scheduling in Intrabay automatic material handling systems

Table 7.1: The membership function of the input variables of weight modification model. Variables

Fuzzy sets (linguistic descriptions)

Lot’s due date satisfaction factor (DDSF)

Bad satisfaction (BS) Normal satisfaction (NS) Good satisfaction (GS) Not heavy (NH) Medium heavy (MH) Very heavy (VH) Long (NL) Medium long (ML) Very long (VL) Very low priority (VLP) Low priority (LP) and High priority (HP), Very high priority (VHP) Most high priority (MHP)

System load factor (SLF)

Lot’s wait time factor (WTF)

The lot’s transport priority

Fuzzy set membership functions (–∞, 0.0, 0.75, 0.85) (0.75, 0.85, 0.95) (0.85, 0.95, 1.0, +∞) (–∞, 0.0, 0.5, 1.0) (0.5, 1.0, 1.5) (1.0, 1.5, 2.0, +∞) (–∞, 0.0, 0.7, 1.0) (0.7, 1.0, 1.3) (1.0, 1.3, 2.0, +∞) (–∞, 0.0, 5.0, 6.0) (5.0, 6.0, 7.0) (6.0, 7.0, 8.0) (7.0, 8.0, 9.0) (8.0, 9.0, 10.0, +∞)

Table 7.2: Fuzzy rule table. Order

Rule ID SLF

1 2 3 4 5 6 7 8 9

VH

MH

NH

DDSF

BS NS GS BS NS GS BS NS GS

WTF NL

ML

VL

VLP VLP VLP VHP HP LP MHP HP LP

VLP VLP VLP VHP VHP LP MHP HP LP

VLP VLP MHP MHP MHP MHP MHP MHP MHP

7.2.4 Hungarian Algorithm-Based Vehicle Dispatching Approach In the previous section, the dynamic priorities of the wafer lots waiting for transportation are computed. In this section, a Hungarian algorithm-based vehicle dispatching approach [8] is developed to allocate wafer lots to automatic vehicles. First, all wafer lots waiting for transportation are sorted by priority in descending order and a certain number of wafer lots with higher priority are selected as the candidate jobs. If the number of wafer lots waiting for transportation Nw is more than the number of

7.2 GDP-Based Intrabay Material Handling Scheduling Method

191

available vehicles Nv, then the first Nv wafer lots with higher priority are selected as the candidate wafer lots. Otherwise, if Nw ≤ Nv, then all Nw wafer lots are selected as candidate wafer lots. Second, the Hungarian algorithm is used to assign the available vehicles to the candidate wafer lots to find the global optimal vehicle dispatching solutions. Third, after the optimal dispatching solutions are found, the available OHT vehicles begin to execute the dispatching solutions according to the greedy vehicle transport policy, which can reduce OHT vehicles’ waste movement effectively. In the Intrabay material handling system, the problem of dispatching available vehicles can be formulated as the following mathematical problem. Assume that there are m wafer lots to be assigned to n available vehicles at any one time. The available vehicles dispatching problem can be formulated as follows: Minimise

n X m X

Cij Xij

(7:16)

Xij = 1, for j = 1, 2, ..., m

(7:17)

i=1 j=1

subject to n X

Xij = 1, for i = 1, 2, ..., n,

i=1

m X j=1

where Xij 2 f0, 1g, and for all i and j, Cij is the travel time of available vehicle j to wafer lot i. The cost matrix C0 is described as 2

C11 C12 6 C21 C22 6 C0 = 4 ... ... Cm1 Cm2

... ... ... ...

3 C1n C2m 7 7 ... 5 Cmn

(7:18)

The solution procedure using the simplified Hungarian method is described as follows. Step1: For the original cost matrix C0 , if m < n, then dummy candidate jobs are added to make the same number. Then the average solution μ is measured as n 1X Cii . μ= n i=1 Step2: The variance σ2ij of each cell in the original cost matrix C0 is calculated to form a new cost matrix C1 . The variance σ2ij is measured as σ2ij = ðCij − μÞ2 . Step 3: For each row in the original cost matrix C1 , the minimum number is subtracted from all numbers. Then a reduced cost matrix C2 is formed. Step 4: For each column in the reduced cost matrix C2 , the minimum number is subtracted from all numbers. Then a reduced cost matrix C3 is formed. Step 5: The minimum number of lines is drawn to cover all zeros in the cost matrix C3 . If this number is equal to n (size of the cost matrix), then stop the process: the optimal solution is obtained. Otherwise, go to Step 4.

192

7 Scheduling in Intrabay automatic material handling systems

Step 6: The minimum uncovered number d is determined. Then it is calculated as follows: (1) d is subtracted from the uncovered numbers; (2) d is added to the numbers covered by two lines and (3) the numbers covered by one line are the same as before. As a result, a new reduced cost matrix C3 is formed. Then go to Step 5.

7.2.5 The Greedy Optimization-Based Vehicle Scheduling Strategy The greedy vehicle scheduling strategy is developed on the basis of the push–pull approach, in which the vehicle is selected according to the states of the current wafer processing machine and the next process machine (i.e., the target processing machine). The procedure of the greedy vehicle scheduling strategy is given in Figure 7.5. When the optimal vehicle dispatching solution is yielded based on the Hungarian algorithm, the OHT vehicle begins to execute the solution. First, the OHT vehicle moves OHT executes optimized dispatching solution

Select the assigned candidate lot

Transport the lot to the Yes destination tool

Is this lot in the stocker’s load port? No

Transport the lot to the Yes stocker

Is this lot’s destination is the stocker

The current transport task is cancelled No

No Are there any lots waiting to enter the current tool? Yes Transport the lot to the Yes destination tool

Are there any idle ports in the lot’s next process tool? No

No

Are there any idle ports in the lot’s next process tool? Yes Transport the lot to the destination tool

Transport the lot to the stocker Figure 7.5: The procedure of the greedy OHT vehicle scheduling strategy in the Intrabay material handling system.

7.2 GDP-Based Intrabay Material Handling Scheduling Method

193

to the original tool where the assigned candidate lot is located. Then a look-ahead mechanism is used to check whether there is any lot waiting to enter the current original tool. If there are some, then a push mechanism-based vehicle transport policy is executed. Otherwise, a pull mechanism-based vehicle transport policy is executed. In the push mechanism-based vehicle transport policy, the state of the next processing tool of the current assigned candidate lot is evaluated. If there is any idle port in the next processing tool, then the current assigned candidate lot is transported to the destination tool’s load port. Otherwise, if there is no idle port in the next processing tool, then the current assigned candidate lot is transported to the stocker. On the other hand, in the pull mechanism-based vehicle transport policy, the state of the next processing tool of the current assigned candidate lot is also evaluated. If there is any idle port in the next processing tool, then the current assigned candidate lot is transported to the destination tool’s load port. Otherwise, if there is no idle port in the next processing tool, the current assigned candidate lot is held in the original tool. Finally, if the optimal vehicle dispatching solution is accomplished or a new empty OHT vehicle arrives, the GDP dispatching policy is reassigned again.

7.2.6 Simulation Experiments In this section, the effectiveness of the proposed approach is demonstrated by using simulation experiments on eM-Plant. The Intrabay data are collected from a 300 mm SWFS. In this study, there are 2 stockers, 28 tools and a 108 m Intrabay rail loop. The average loading/unloading operation time of the OHT vehicles is 16 seconds. The speed of the OHT vehicles is 1.8 m/s. The number of OHTs can be 5, 6 and 7 in the Intrabay system. The system loading ratios can be 90%, 100% and 110% of the Intrabay system. Each Intrabay experiment was run for 15 days with a warm-up of two days. The proposed approach is compared with the traditional first encounter first served (FEFS) rule and the highest priority lot first rule (HP) in order to evaluate its effectiveness. The experimental results of the GDP approach are summarized in Tables 7.3 and 7.4. The results demonstrate that the GDP approach outperforms the FEFS rule- and HP rule-based dispatching approaches. A desirability function DðXÞ is used to compute the comprehensive performance indexes in order to further verify the effectiveness of the proposed approach. The desirability function is a geometric mean of all performance indexes:

D=

n Y

!1=n di

(7:19)

i=1

where di is the individual desirability function of the performance index and n is the number of performance indexes of the automatic material handling system and the SWFS. In this study, all performance indexes are treated as having the same weight.

100%

90%

System loading

Scenario

7

6

5

7

6

5

OHT number

FEFS HP GDP FEFS HP GDP FEFS HP GDP FEFS HP GDP FEFS HP GDP FEFS HP GDP

Dispatching policy

64,218 59,479 58,478 63,310 59,481 58,478 62,926 58,482 58,481 64,868 64,868 64,849 64,868 64,871 64,857 64,866 64,871 64,853

Mean value 8.9 1.7 / 7.6 1.6 / 7.0 0 / 0 0 / 0 0 / 0 0 /

Improved (%)

Movement (lot)

Table 7.3: Comparison of the GDP approach and FEFS rule.

117.0 125.5 112.3 106.7 105.4 101.9 102.0 99.3 97.4 120.4 121.7 113.3 109.5 108.6 106.8 104.4 105.0 102.0

Mean value 4.0 10.5 / 4.4 3.3 / 4.5 1.9 / 5.9 6.9 / 2.5 1.7 / 2.3 2.9 /

Improved (%)

Delivery time (s)

50.5 49.4 49.4 51.1 50.4 50.3 51.6 51.0 50.7 51.3 51.0 50.5 51.9 51.3 51.0 52.8 52.5 50.7

Mean value 2.2 0.0 / 1.6 0.2 / 1.7 0.6 / 1.6 1.0 / 1.7 0.6 / 4.0 3.4 /

Improved (%)

Transport time (s)

49.51 44.74 44.20 41.15 38.48 37.37 35.45 32.51 32.23 50.74 50.43 50.08 42.79 42.37 42.22 37.05 37.42 35.96

Mean value

(continued )

10.7 1.2 / 9.2 2.9 / 9.1 0.9 / 1.3 0.7 / 2.9 3.9 / 0.3 0.4 /

Improved (%)

OHT utilization (%)

194 7 Scheduling in Intrabay automatic material handling systems

Total average

110%

Table 7.3: (continued )

7

6

5

FEFS HP GDP FEFS HP GDP FEFS HP GDP FEFS HP GDP

69,905 70,280 69,529 69,729 69,953 69,598 70,031 69,627 69,579 66,080 64,657 64,300

0.6 1.1 / 0.2 0.6 / 0.6 0.1 / 2.7 0.6 /

142.6 151.0 135.0 125.9 133.3 124.0 127.6 123.5 112.4 117.3 119.3 111.7

5.3 10.5 / 1.5 7.0 / 11.9 9.0 / 4.8 6.4 /

53.1 53.0 53.0 53.3 53.4 53.1 55.1 54.7 54.1 52.3 51.9 51.4

0.2 0 / 0.4 0.6 / 0.2 1.1 / 1.7 1.0 /

56.66 56.95 56.25 47.26 47.36 47.17 41.44 41.29 41.23 44.67 43.51 42.97

0.7 1.2 / 0.2 0.4 / 0.5 0.1 / 3.8 1.2 /

7.2 GDP-Based Intrabay Material Handling Scheduling Method

195

100

90

System loading (%)

Scenario

5 FEFS HP GDP 6 FEFS HP GDP 7 FEFS HP GDP 5 FEFS HP GDP 6 FEFS HP GDP 7 FEFS HP GDP

OHT number

Dispatching policy

2,657.8 2,622.6 2,553.7 2,561.8 2,503.0 2,488.4 2,523.7 2,467.7 2,458.9 3,178.4 3,165.0 3,108.4 3,116.9 3,096.5 3,084.2 3,081.5 3,079.2 3,058.8

Mean value 3.9 2.6 / 2.8 0.6 / 2.6 0.4 / 2.2 1.8 / 1.0 0.4 / 0.7 0.6 /

Improved (%)

Cycle time (s)

Table 7.4: Comparison of the GDP approach and FEFS rule.

13,658 13,657 13,658 13,658 13,659 13,659 13,657 13,659 13,659 15,152 15,152 15,152 15,152 15,152 15,152 15,152 15,152 15,152

Mean value 0 0 / 0 0 / 0 0 / 0 0 / 0 0 / 0 0 /

Improved (%)

Throughput (lot)

98.63 98.72 100.0 98.63 98.72 100.0 98.63 98.72 100.0 98.42 98.52 100.0 98.42 98.52 100.0 98.42 98.52 100.0

Mean value 1.4 1.3 / 1.4 1.3 / 1.4 1.3 / 1.6 1.5 / 1.6 1.5 / 1.6 1.5 /

Improved (%)

Due date satisfaction (%)

30 29 29 29 28 28 28 28 28 39 39 38 38 38 38 38 38 38

Mean value

(continued )

3.3 0 / 3.4 0 / 0 0 / 2.6 2.6 / 0 0 / 0 0 /

Improved (%)

WIP in Intrabay (lot)

196 7 Scheduling in Intrabay automatic material handling systems

Total Average

110

Table 7.4: (continued )

5 FEFS HP GDP 6 FEFS HP GDP 7 FEFS HP GDP FEFS HP GDP

25,496.0 25,405.0 18,988.0 24,824.0 24,651.0 18,458.7 24,381.0 24,423.0 18,296.0 10,202.3 10,157.0 8,055.0

25.5 25.2 / 25.6 25.1 / 24.9 25.0 / 21.0 20.7 /

16,100 16,206 16,263 16,124 16,100 16,276 16,108 16,100 16,278 14,973 14,982 15,028

1.0 0.4 / 0.9 1.1 / 1.1 1.1 / 0.4 0.3 /

94.34 94.67 99.29 94.29 94.58 97.15 94.27 94.63 99.58 97.12 97.29 88.56

5.2 4.9 / 3.0 2.7 / 5.6 5.2 / 8.8 9.0 /

340 290 261 331 342 254 339 339 252 135 130 107

23.2 10 / 23.3 25.6 / 25.6 25.7 / 20.7 17.7 /

7.2 GDP-Based Intrabay Material Handling Scheduling Method

197

198

7 Scheduling in Intrabay automatic material handling systems

Using the desirability function DðXÞ, the desirability values of the FEFS rule, HP rule and proposed GDP approach in different scenarios are calculated, as shown in Figure 7.6. A statistical test of desirability values using analysis of variance (ANOVA) is executed, as shown in Table 7.5. The F value of 26.72 in Tab. 7.5 indicates that there are significant differences between the FEFS rule and the proposed GDP approach. The F value of 16.22 in Table 7.5 means that there are significant differences between the HP rule and the proposed approach. The comparison of the desirability values of the proposed GDP approach, FEFS rule and HP rule in Figure 7.6 illustrates that the proposed GDP approach clearly outperforms the other two rules in terms of comprehensive performance. (In Figure 7.6, L indicates system loading ratios; e.g., L90 means 90% loadings in the Intrabay system. V indicates the OHT vehicle number; e.g., V5 means five OHT vehicles in the Intrabay system.) In all experimental scenarios, the computation time of the proposed GDP approach is more than the FEFS rule- and HP rule-based OHT vehicle dispatching

Table 7.5: ANOVA of the FEFS rule and the HP rule with GDP dispatching approach. Compared rules Source FEFS and GDP

HP and GDP

Columns Error Total Columns Error Total

Sum of squares (SS) 0.20895 0.12512 0.33408 0.12366 0.122 0.24566

0.8

1 16 17 1 16 17

0.20895 0.00782 – 0.12366 0.00762 –

F-value

Prob.>F

26.72 – – 16.22 – –

< Mind ^ d½i = d½i > otherwise > > : Mind

(8:11)

(8:12)

8.2.3.3 Genetic operators in PMOGAs 1.

Path encoding A real-coded manner is adopted by the path encoding process. Denote G = ðN, EÞ, in which N = fn0 , n1 , ..., nr g is the node set; E  ðN × NÞ represents a set of directional links between nodes; n, n′ 2 L indicates the link from node n to n' Cpm , AdjðnÞ = fm : ðn, mÞ 2 Lg represents the adjacent node set of node n. First, all the links of consecutive nodes in the system are numbered. Assume that there are r nodes, m links, then the serial numbers are assigned in accordance with the following manner: ðn1 , n2 Þ = 1, ðn2 , n3 Þ = 2, ..., ðnr − 1 , nr Þ = m

(8:13)

Denote path p = fn0 , n1 , ..., nk g as a sequence from node n0 to nk . This node sequence corresponds to a sequence of links, and each path can be represented by a string of serial numbers from 1 to m. Since the walking route of a vehicle is different, the variable length chromosome is used to represent individuals.

8.2 PMOGA-Based Integrated Scheduling Method in AMHSs

219

If there are k OHTs, then a multi-parameter cascade coding is used to represent a chromosome, each vehicle occupies xi digit (i = 1, 2, . . .. . ., k), as follow: Chromosome = b11 b12 ...b1l1 jb21 b21 ...b2l2 j...jbk1 bk2 ...bklk The total length of the string is

k P

(8:14)

li .

i=1

2.

Population initialization The initial population has a greater impact on the results. Artificial selection method is not applicable to the present problem, and randomly generated rule will generate a lot of unavailable paths. Solomon [5] proposed a heuristic algorithm for VPR to generate a good individual, and then generate some individuals in the neighborhood of the individual, which account for a certain proportion and the rest of the individuals are randomly generated. Eppstein proposed a K shortest path algorithm that can be used to generate an initialization path. In addition, there are network flow methods and some point-to-point shortest path generation algorithms. Since this part of the calculation is off-line calculation, the time complexity of the algorithm would not be taken into consideration. Dijkastra-Floyd algorithm [7] is used to generate the initial optimal path, and part of solutions is selected as the initial population. In the population initialization procedure, the shortest path from any vertex in graph G = (N, L) to all other vertices is found to calculate the initial optimal path. The weighted adjacency matrix of G is presented as follow: 0

0 B d21 B B D=B ∞ B @... ∞

d12 0 d32 ... ∞

∞ d23 0 ... ...

... ... ... ... dn, n − 1

1 ∞ ∞C C C ∞C C ...A

(8:15)

0

In addition to the subscript elements and diagonal elements in matrix D, the remaining elements are infinite. The shortest distance and optimal path set DFSet between all vertex pairs in the Interbay material handling system is obtained by using Dijkastra–Floyd algorithm. The remaining paths are generated by using a random algorithm. In each real-time scheduling trigger, the appropriate path is selected from DFSet as the initial path according to the OHT starting position corresponding to the node. 3.

Select operator Random tournament selection operator [8] is adopted in this algorithm. This selection method has both random and deterministic characteristics, and its calculation efficiency is very high. Since PMOGA is a multi-objective algorithm, the steps of its selection operator are presented as follows:

220

8 Integrated scheduling in AMHSs

(1) n (n = 2) individuals are randomly selected from the population for comparison, and the individuals with the highest rank are inherited to the next generation. If the sort is the same, the individuals with smaller density are selected. (2) The above process is repeated N times, and N individuals are generated for the next generation. 4. Crossover operator Due to the special nature of OHT movement, the single point or multipoint crossover operator will produce discontinuous infeasible path, which is not conducive to the evolutionary convergence of the algorithm. Therefore, a hybrid node crossover operator is used in PMOGA. For two randomly selected individuals, only the points with the exactly same serial number are selected for crossover operation. When there is more than one coincidence node, a node is randomly selected. If there is less than one node, no crossover operation will be performed. This ensures that no intermittent paths are generated. The procedure is illustrated in Figure 8.6. The basic steps of the crossover operator are presented as follows: (1) Two paths R1 and R2 with a common node are randomly selected from the population in accordance with crossover probability Pc; (2) A node from all the common nodes is selected as a crossover point; (3) All sub-path (node sequence) positions are exchanged after the intersection is crossed 5.

mutation operator, insert operator, and delete operator. The mutation operation of a PMOGA consists of three operations. (1) Randomly remove a serial number from the chromosome (excluding the start and end numbers); (2) Randomly select a node from the chromosome to insert a new serial number; (3) Randomly select a serial number in the individual and replace it with a new serial number. Since these three ways are likely to produce

Cross node ...

R1 R2

Crossover

Possible cross node Selected cross node

...

Other path node Cross node

Figure 8.6: Schematic diagram of the path crossover.

8.2 PMOGA-Based Integrated Scheduling Method in AMHSs

221

intermittent path, two special operators (an insert operator and a delete operator) are added. For example, for path R, the mutation operation, insertion operation and deletion operation are presented as follow: – R = (13, 14, 15, 7, 8, 9, 17) – Mutation: R = (13, 14, 15, 12, 8, 9, 17) – Insertion: R = (13, 14, 15, [16, 17, 21], 12, [13, 18], 8, 9, 17) – Deletion: R = (13, 14, 15, 16, 17) 8.2.3.4 Fitness evaluation As the previous presentation, if E  ðN × NÞ contains m links, then a vector is used to represent the status of the rail at scheduling time t:   s = S1 ðtÞ, S1 ðtÞ, ..., Sm ðtÞ Suppose that  l

S =

1 0

if link l is blocked otherwise

(8:16)

If there is at least one busy OHT on the link l, it is called blocked. For any link l, the status vector fSl ðtÞ, t = 0, 1, ...g related to status Sl ðtÞ at time t = 0, 1, … is a Markov chain, and any two status vectors fSl1 ðtÞ, t = 0, 1, ...g and fSl2 ðtÞ, t = 0, 1, ...g are Independent [9], in which l1 l2 . The status transition matrix is:  Pt,l t + 1 =

1 − αlt 0

αlt 1

 (8:17)

wherein αlt represents the probability that the directed link l is blocked before time t, and is idle after time t, e.g., Sl ð0Þ = 1, Sl ð1Þ = 1, . . . , Sl ðtÞ = 1, Sl ðt + 1Þ = 0. The status transition matrix shows probability αlt that any directed link l changes from blocked to idle at time t, as shown in Figure 8.7. Three factors including the average OHT delivery time, longest OHT delivery time and average processing cycle of a workpiece are considered in the fitness evaluation process. The smaller the fitness value, the better the individual.

1 – αtl

Blocked (1)

αtl

1

Idle (0)

Figure 8.7: Sketch map of the state transition characteristics of link l.

222

8 Integrated scheduling in AMHSs

3 4

5

1 2

3

2

4

1

5

Figure 8.8: Exchange operator in the local search.

8.2.3.5 Local search strategy A λ-interchange local search method is adopted in the local search strategy. Since the processed individuals in PMOGAs are paths, the local search process improves the quality of the solution by exchanging nodes in different paths. For example, for a given path (Rp, Rq), the exchange operator (1, 2) represents the replacement of one node from path Rp to path Rq, and one node is changed from path Rq to path Rp., as shown in Figure 8.8. The FS strategy is used to select switching nodes. Due to the difference between multi-objective and single-objective optimization problems, the solution S’ that leads to the cost reduction is an optimal solution to dominate the current solution. The mixed local search strategy can significantly improve the quality of the genetic algorithm, but on the other hand, it will increase the computational complexity. Therefore, it only assigns a higher local search probability to the non-dominated individuals at the forefront of Pareto (the highest rank). In the experiment, the initial setting probability of the local search process is 10% and the local search probability of remaining individuals is 1%. 8.2.3.6 PMOGA execution process The overall execution process of a PMOGA is illustrated in Figure 8.9. The initial population is divided into several sub-populations according to the number of processors, and each sub-population is concurrently running on the respective processors. After a certain evolutionary iteration, each sub-population will exchange a number of individuals in accordance with the migration topology and the migration strategy. In the process of merging sub-populations, in order to maintain the new final population size N (the size of the sub-population Popi), when Popt i + Fi > N, select N-Popt i + Fi individuals with the smallest aggregation distance from Fi to enter the new population. When the algorithm finally converges, an optimal Pareto solution set is obtained, and the individual with the largest clustering distance is chosen as the final optimal solution from the optimal solution set.

8.2.4 Simulation Experiments 8.2.4.1 Computational Efficiency Analysis The PMOGA parallelization scheme proposed in this chapter is developed on the basis of C ++ and OpenMP, and the basic parameters of GA and parallel parameter

8.2 PMOGA-Based Integrated Scheduling Method in AMHSs

223

Start Master node initializes system date information, create child node, transport data Dijkastra-Floyd shortest path algorithm + stochastic generate initial population Generate subpopulation Pop1

Generate subpopulation PopN

Merge current population and former population t t t–1 PN1 = Pop1 + Pop1

Quick sort non-dominated set construction method

Merge current population and former population t t t–1 PN1 = Pop1 + Pop1

...

t

t

Run quick Pareto sort on PN1

Select individuals and place them in the new t+1 population Pop1 until the new population is filled

Run quick Pareto sort on PN1

...

t+1

t+1

Pop1 = ∅, i = 1

Pop1 = ∅, i = 1 i = i+1

i = i+1 t+1 | Pop1 |

+ | Fi | < N

N

t+1

| Pop1 | + | Fi | < N

Apply aggregation calculation on Fi

Fore-end Fi of ith Pareto Definite partial order according to dominace relationship and individual density

Apply aggregation calculation on Fi

...

Merge populations t+1 t+1 Pop1 = Pop1 ⋃ Fi

Merge populations t+1 t+1 Pop1 = Pop1 ⋃ Fi

Further distinguish Fi according to the partial order

Further distinguish Fi according to the partial order

Obtain new population t+1 t+1 Pop1 = Pop1 ⋃ Fi[1→(N–| Pt+1 |)]

Obtain new population t+1 t+1 Pop1 = Pop1 ⋃ Fi[1→(N–| Pt+1 |)]

Migration topology/ strategy

Run selection, interaction, mutation, insertion, delete operator

Migration opertion

Local search strategy t = t+1 Meet migration cycle TM?

Run selection, interaction, mutation, insertion, delete operator Local search strategy Y

Y

t = t+1 Meet migration cycle TM?

N N

Meet termination condition? Y

N N

Meet termination condition? Y

Output optimal non-dominated solution set, end

Figure 8.9: General flowchart of PMOGA.

N

224

8 Integrated scheduling in AMHSs

Table 8.1: The basic parameters of a GA and parallel parameter settings. Name

Value

The maximum length of chromosome Iteration number Population size Objective number CPU number

200 200–500 200–400 3 4

Name

Value

Crossover probability Mutation probability λ-Interchange local Search probability d-Value

0.3 0.1 0.1–0.2 2

Table 8.2: The basic parameters of a GA and parallel parameter settings. Index 1 2 3 4 5 6

Objective number

Population size

Iteration number

λ

CPU number

CPU time (ms)

3 3 3 3 3 3

200 200 200 400 400 400

200 200 500 200 500 500

0.1 0.1 0.1 0.1 0.1 0.2

1 4 4 4 4 8

1,700 540 540 911 2,139 1,218

settings are shown in Tables 8.1 and 8.2. (The same experiments are repeated three times to obtain the average values.) 8.2.4.2 Convergence Analysis Since there is no known optimal solution set PFtrue for the multi-objective problem currently studied, the approaching degree evaluation method [10] is used, in which the approximation is measured by calculating the minimum distance of the solution set to the reference set. The smaller the degree of deviation, the higher the quality of solution. The optimal Pareto set of solutions and the convergence curves of the three objectives and the two objectives (only considering objectives 1 and 2) are shown in Figures 8.10 and 8.11, respectively. 8.2.4.3 Comprehensive analysis The integrated scheduling method based on PMOGA proposed in this chapter is compared with the previous scheduling method in three indexes (average delivery time, maximum delivery time and average workpiece processing cycle). The basic parameters and process parameters are the same as given in Section 8.2. Three release rates (3 cards/2.5 h, 3 cards/2.75 h and 3 cards/3 h), and three OHT numbers (8, 10, 12) are implemented. The total number of simulation scenes is 3 × 3 = 9, and the experiment is repeated once for each scene. The results obtained are shown in Figures 8.12 to 8.15.

8.2 PMOGA-Based Integrated Scheduling Method in AMHSs

225

3 Fitness value f3

2.5 2 1.5 1 0.5 0 180

280

160 (a)

Fitness value f2

140

240 120

220

300

260 Fitness value f1

180 170 160 150 140 130 120 220

230

240

250

260

270

280

290

300

(b) Figure 8.10: Optimal Pareto solution sets: (a) optimal Pareto solution distribution when there are three objectives and (b) optimal Pareto solution distribution when there are two objectives.

There are no significant differences for these methods in the average delivery time. However, the PMOGA-based scheduling method can significantly reduce the maximum delivery time of scheduling tasks, which is important for an Interbay material handling scheduling problem with time constraints. For the average processing cycle, the PMOGA-based method has no significant difference compared with the GPCRG-R1 and MHAFLC methods. For the distribution of processing time (Figure 8.16), since the multi-objective optimization algorithm proposed in this chapter balances multiple optimization objectives, the average processing time is in the vicinity of the mean, which has a small mean difference.

226

8 Integrated scheduling in AMHSs

0.7 Convergence curve 1 Convergence curve 2 Convergence curve 3

Approaching degree (CP(t))

0.6 0.5 0.4 0.3 0.2 0.1 0

0

10

20

30

40

50

60

70

80

90

100

Iterations

The average delivery time (s)

Figure 8.11: Convergence process of optimal Pareto solution sets.

800 700 PMOGA

600

MHAFLC

500

GPCRG-R1

400

EDD

300

FEFS

200

RLWT

100

na

rio -9 en a

rio -8

Sc

-7 Sc e

Sc

en

ar

io

rio -6

-5

na Sc e

ar en

Sc

Sc

en

ar

io

io

-4

-3 io

-2

en ar

io

Sc

en ar

Sc

Sc

en a

rio -1

0

Figure 8.12: Comparison on average delivery time.

8.3 GARL-Based Complex Heuristic Integrated Scheduling Method in AMHSs In this chapter, on the basis of GP-based complex heuristic rule constructing method presented in Chapter 6, a GARL-based complex heuristic integrated scheduling method is proposed to complete the integrated scheduling process for the wafer processing, vehicle assignment and path planning by using path planning methods and specific dispatch rules. The integrated scheduling mathematical model

227

8.3 GARL-Based Complex Heuristic Integrated Scheduling Method in AMHSs

The longest delivery time (s)

1,200 1,000

PMOGA MHAFLC

800

GPCRG-1

600

EDD FEFS

400

RLWT 200

Sc en

ar io -6 Sc en ar io -7 Sc en ar io -8 Sc en ar io -9

ar io -5

ar io -4

Sc en

Sc en

ar io -3

ar io -2

Sc en

Sc en

Sc en

ar io -1

0

Figure 8.13: Comparison on longest delivery time.

Average processing cycle (s)

3,000,000 2,500,000 PMOGA 2,000,000

MHAFLC GPCRG

1,500,000

EDD FEFS

1,000,000

RLWT 500,000

na Sc e

en

ar

io

rio

-9

-8

-7 rio Sc

na Sc e

en

ar

io

rio Sc

en a

io Sc

en ar

Sc

-6

-5

-4

-3 rio na

io Sc e

en ar Sc

Sc

en ar

io

-2

-1

0

Figure 8.14: Comparison on average processing cycle.

established in Section 8.2.1 is adopted to obtain the optimal Pareto solution for minimizing the average vehicle delivery time and minimizing the production cost by searching for reasonable vehicle dispatch rules and vehicle paths.

8.3.1 Basic Rules for Automated Material Handling System Scheduling The GARL-based integrated scheduling method consists of two parts: vehicle assignment composite heuristic rule generation and vehicle path planning. The vehicle assignment composite rules are developed on the basis of the assignment rules shown in Table

228

8 Integrated scheduling in AMHSs

Table 8.3: Basic assignment rules in AMHSs. Categories

Index

Description

S1

S11 S12 S13 S14 S21 S22 S23 S24 S31 S32 S33 S34

First in first out Shortest remaining time priority Critical ratio priority The earliest plan priority First in first serve Longest waiting part priority Shortest handling distance priority Longest handling distance priority Nearest idle vehicle priority Farthest idle vehicle priority Longest idle time vehicle priority Lowest utilized vehicle priority

S2

S3

3,000

Amount (lot)

2,500

>35000 30001–35000

2,000

25001–30000 1,500

20001–25000 15001–20000

1,000

10000–150000 L1+L2?

L2>0?

N Wafer lot priority

L2 = L2+1

N

N K1 = K1–1 K2 = K2+1

Y

Y

S1 compound rule

L3 = L–L1–L2

K2>0?

PL3 = Φ?

Y End

S2 compound rule

Wafer lot priority

L1 = L1+1

N

Y

S3 compound rule

L1 = L1+1

N S1 compound rule

K1 = K1+1 K2 = K2–1 R L1 path selection End

Figure 8.17: Chromosome decoding of OHT dispatching and path planning.

is 3 × 3 = 9, and the experiment is repeated once for each scene. The results obtained are shown in Figures 8.18 to 8.19. The algorithm proposed in this chapter can significantly reduce the handling cost and the average response time during operation in AMHSs. On one hand, these results are benefited from the proposed integrated scheduling method of the vehicle assignment compound rule generation and vehicle path selection. On the other hand,

8.3 GARL-Based Complex Heuristic Integrated Scheduling Method in AMHSs

231

Table 8.4: Basic GA parameters and parallel setting parameters. Parameter name Maximum chromosome length Evolutionary iteration number Population size

Parameter value

Parameter name

Parameter value

200

Crossover probability

0.3

200

Mutation probability

0.1

400

λ-Interchange local search probability

0.1–0.2

Table 8.5: Processing technology and key equipment data. Warehouse number

Bay number

Corresponding process

Number of critical equipment

Processing time (min)

Equipment processing capacity (lot)

Stocker-1

Bay-1

Stocker-2

Bay-2

Wash WIS Lithography Sputtering 1 Electrical Testing Sputtering 2 Sputtering 3 Chemical vapor deposition Wash WCW Chemical vapor deposition Automatic appearance Wash WCW Automatic appearance Sputtering 2 Sputtering 3 Lithography Stripping Automatic appearance Repair (excision) Wash WCW Repair (connection)

3 4 3 3

20 73 35 44

2 3 2 1

1 1 5

28 26 43

2 2 2

3 5

19 41

2 3

2

48

1

4 1

21 48

3 1

1 2 5 4 3

28 26 34 24 114

2 2 2 2 2

5

40

1

1 2

22 10

2 1

Stocker-3

Bay-3

Stocker-4

Bay-4

Stocker-5

Bay-5

Stocker-6

Bay-6

Stocker-7

Bay-7

(continued )

232

8 Integrated scheduling in AMHSs

Table 8.5: (continued ) Warehouse number

Bay number

Corresponding process

Stocker-8

Bay-8

Stocker-9

Bay-9

Dry etching DEC Dry etching DEI Dry etching DCH Automatic appearance Array check Automatic appearance Repair (connection) Appearance inspection Automatic appearance Repair (excision) Wash WCW Wet etching WEI Annealing Wash WCW Wet etching WEI Wet etching WEG Stripping Wet etching WED Dry etching DEI Wet etching WEG Wet etching WED Dry etching DEI Dry etching DCH Dry etching DEC Wash WCW Dry etching DEI Lithography Dry etching DEC

Stocker-10

Bay-10

Stocker-11

Bay-11

Stocker-12

Bay-12

Stocker-13

Stocker-14

Bay-13

Bay-14

Number of critical equipment

Processing time (min)

Equipment processing capacity (lot)

1 1 1 2

46 41 53 48

2 2 2 1

2 3

91 48

2 1

3

10

2

5

29

1

3

114

2

3

34

1

1 1 2 1 2 1

20 41 101 18 41 28

2 3 5 2 3 2

3 1

24 34

2 2

1 2

41 28

2 2

4

34

2

1 2 2 11 1 7 5

41 53 46 20 41 79 46

2 3 2 1 2 4 2

8.4 Conclusion

233

3,000,000 2,500,000 2,000,000

GA EDD

1,500,000

FEFS RLWT

1,000,000 500,000

Sc en

ar io -1 Sc en ar io -2 Sc en ar io -3 Sc en ar io -4 Sc en ar io -5 Sc en ar io -6 Sc en ar io -7 Sc en ar io -8 Sc en ar io -9

0

Figure 8.18: Comparison on average handling cost.

1,200 1,000 800

GA EDD

600

FEFS RLWT

400 200

rio -9 na

rio -8

Sc e

-7 io

na Sc e

Sc

en

ar

io

-6

-5 io

en ar Sc

-4 ar

io

Sc

en ar

-3 io en Sc

Sc

en ar

-2 io ar

en Sc

Sc e

na

rio -1

0

Figure 8.19: Comparison on average response time.

they are also benefited from the Pareto domination for multi-objective problems to find the optimal Pareto solution.

8.4 Conclusion In this chapter, the integrated scheduling problems of Interbay and Intrabay material handling processes in wafer fabrication systems have been investigated and their corresponding scheduling methods have been introduced. The integrated scheduling

234

8 Integrated scheduling in AMHSs

method based on parallel multi-objective genetic algorithms and the integrated scheduling method based on multi-agent cooperation have been presented in detail. The methods and algorithms have been verified by using the data from a Shanghai wafer fabrication manufacturer.

References [1]

Le-Anh, T., De-Koster, M.B.M. A review of design and control of automated guided vehicle systems. European Journal of Operational Research, 2006, 171: 1–23. [2] Vis, I.F.A. Survey of research in the design and control of automated guided vehicles systems. European Journal of Operational Research, 2006, 170: 677–709. [3] Yin-Bin, S. Intelligent Scheduling Wafer Fabrication Interbay Material Handling System. Shanghai Jiaotong University, 2011. [4] Jensen, M.T. Reducing the run-time complexity of multiobjective EAs: The NSGA-II and other algorithms. IEEE Transactions on Evolutionary Computation, 2003, 7(5): 503–515. [5] Solomon, M.M. Algorithms for vehicle routing and scheduling problems with time window constraints. Operations Research, 2010, 35(35): 254–266. [6] Eppstein, D. Finding the k shortest Paths in a network. Management Science, 1971, 19(17):712-716. [7] Cormen, T.H., Leiserson, C.E., Rivest, R.L., et al. Introduction to Algorithms (Second Edition). Cambridge: The MIT Press, Mcgraw-Hin, 2001, pp. 1297–1305. [8] Liao, D.Y., Wang, C.N. Differentiated preemptive dispatching for automatic materials handling services in 300 mm semiconductor foundry. The International Journal of Advanced Manufacturing Technology, 2006, 29(9): 890–896. [9] Osman, I.H., Metastrategy simulated annealing and tabu search algorithms for the vehicle routing problem. Annals of Operations Research, 1993, 41(4): 421–451. [10] Zhen-Cai, C., Qi-Di, W., Fei, Q., Zun-Tong, W. Advances in semiconductor wafer fabrication scheduling based on genetic algorithm. Journal of Tongji University, 2008, 36(1): 97–102.

9 Scheduling Performance Evaluation of Automated Material Handling Systems in SWFSs Over the last few years, performance evaluation and analysis of the AMHS have received considerable attention. This chapter investigates the modeling process of the AMHS in order to evaluate the performances of various scheduling methods. Since the production scheduling is within a real-time environment, the evaluation model is required to be not only effective but also efficient. An agent-oriented and knowledge-based colored timed Petri net (AOKCTPN)-based approach is therefore proposed.

9.1 Modeling Requirements of Scheduling Performance Evaluation in the AMHS Traditional modeling approaches for manufacturing systems include queuing networks, perturbation analysis, Markov chain, Petri net (PN) and so on. Due to the special characteristics of a SWFS, the PN as an analytical tool for discrete event dynamic systems has become the most widely adopted method for modeling semiconductor manufacturing systems and analyzing the schedule performances. In recent years, there have been many reports of PN and extended PN applications in modeling SWFSs. Object-oriented Petri net (OOPN) methodologies were widely investigated in order to reduce the model scale and improve its re-configurability and re-usability. Zhang et al. [1] proposed an extended OOPN modeling approach to build up a simulation platform for real-time scheduling of SWFS. Liu et al. [2, 3] proposed an extended object-oriented Petri nets (EOPNs) for effectively modeling SWFSs. The resulting model validates that the EOPNs may cope well with complex SWFSs modeling. Wang and Wu [4] proposed an object-oriented hybrid PN model (HOPN) approach for semiconductor manufacturing lines, which provides a good foundation for studying scheduling policies. Zhou et al. [5] proposed an object-oriented and extended hybrid PNs to describe a semiconductor wafer fabrication flow. The modeling results indicated that the proposed modeling approach can deal with the dynamic modeling of a complex photolithography process effectively. However, most of the approaches proposed in above researches are applied for SWFSs; meanwhile few studies can be found for modeling AMHSs. On the other hand, the OOPN approach can only support for the predefined control flow, not for distributed dynamic route control in AMHSs. For the distributed environment, agent-oriented PN methodologies get increasing attention in semiconductor manufacturing systems. Zhang et al. [6] and Zhai et al. [7] constructed an AOCTPN for SWFSs, which reduced the complexity and https://doi.org/10.1515/9783110487473-009

236

9 Scheduling Performance Evaluation

improved the reusability of the model, and could be used to describe the coordination scheduling and controlling between agents. However, in these agent-oriented PN approaches, the information and knowledge of scheduling and controlling process of SWFSs and AMHSs cannot be easily described. Therefore, the established models have poor performances in supporting the decision-making. Based on the above brief literature review, some conclusions are obtained and listed as follows: 1. Although the PN has been widely used in modeling and performance analyzing for SWFSs and AMHSs, most of these models are not suitable for large-scale and complex SWFSs and AMHSs, especially for the 300 mm wafer fab. In addition, such PN-based models are highly system dependent and lack re-configurability and re-usability. Therefore, the adaptability of the model for changing market and dynamic environment cannot be guaranteed. 2. The OOPN-based modeling approach can deal with the dynamic modeling for a complex manufacturing system and has been used to model and analyse SWFSs. However, few researches in this field can be found for AMHSs. At the same time, traditional OOPN approaches can only support the predefined control flow, which may not be applied for distributed dynamic route control in AMHSs. 3. In order to support decision-making such as scheduling and controlling, the model for AMHSs should be capable of describing the status information of AGV and furthermore the knowledge of scheduling and controlling process of AMHSs. However most of the PN-based approaches cannot provide such information and knowledge and therefore are not easy to be integrated with the decision-making process. In order to fill the above research gaps, the chapter introduces a modeling approach based on an AOKCTPN to evaluate the performances of various scheduling methods. An AOKCTPN is an object-oriented approach, which can achieve great reusability, and adopts knowledge-based decision-making methods for places and transitions in PNs, which can improve the decision-making ability and adaptability.

9.2 A Modeling Approach Based on an Agent-Oriented Knowledgeable Colored and Timed Petri Net (AOKCTPN) The method based on the simulation model and the method based on the PN model are less reusable and inefficient in evaluating the scheduling performances of an AMHS. In this chapter, the advantages of the PN model in dealing with dynamic discrete events and complex parallel systems are adopted. At the same time, the advantages of the high-level OOPN model in reducing the size and reusability of the problem are combined. Therefore, this chapter proposes a reusable and modular

9.2 Agent-Oriented Knowledgeable Colored and Timed Petri Net

237

scheduling performance evaluation modeling method: agent-oriented knowledgeable colored and timed Petri net (AOKCTPN).

9.2.1 Definition of an AOKCTPN Model 1.

An AOKCTPN method An AOKCTPN is developed on the basis of an agent-oriented PN modeling theory. By improving the agent-oriented PN modeling theory and expanding the knowledge description of the place and transition based on the object model, the complexity of large-scale SWFS modeling is effectively solved. According to the real SWFS hierarchical architecture, the AOKCTPN model is built up in a hierarchical form in order to make the established system model more concise and accurate. An AOKCTPN can be defined as AOKCTPN = fAPS, TC, C, D, A, KB, Mi g

where 1.

2.

3.

APS = APS = fAP1 , . . . , APi , . . . APn g, which is the set of agents in the model. APS includes basic agent place set and composite agent place set, which is denoted by APS = fLAPSgU fLCAPSg. LAPS = fLAP1 , LAP2 , . . . , LAPi , . . . , LAPw g, which corresponds to a set of basic agent places in the model, where LAPi is the ith agent place in the set. LAPi represent the physical resource agents (such as processing machine, transportation rail, shortcut, turntable and stocker) and logical resource agents (such as releasing, dispatching, routing) in an AMHS. LCAPS = fLCAP1 , LCAP2 , . . . , LCAPi , . . . , LCAPw g, which corresponds to a set of composite agent places in the model, where LCAPi is the ith composite agent place in the set, which is denoted by LCAPi = fLAPS, TC, C, D, I, O, KB, M0 g.   TC = TC1 , TC2 , . . . TCj , . . . TCm , which corresponds to a limited set of communication transitions, where TCj is the jth communication transition that provides communication among basic agent places and composite agent places. In addition, LAPi ∪ TCj ≠ Φ, LCAPi ∪ TCj ≠ Φ, LAPi ∩ TCj = Φ, and LCAPi ∩ TCj = Φ.     C = fCðAP1 Þ, . . . , CðAPi Þ, . . . , CðAPn ÞgU CðTC1 Þ, . . . , C TCj , . . . , CðTCm Þ , which corresponds to color set of places and communication transitions.   CðAPi Þ = ai1 , ai2 , . . . aiμi , , which is a color set of agent places APi , where o   n μi = jCðAPi Þjði = 1, 2, . . . , nÞ. C TCj = bj1 , bj2 , . . . bjμj , , which is a color   set of agent places TCj , where μj = CðAPj Þðj = 1, 2, . . . , mÞ.

238

9 Scheduling Performance Evaluation

4. D corresponds to time delay set of places and transitions. DðAPi Þ = fdil , . . . , djh . . . , diu g, μi = jDðAPi Þj = jCðAPi Þj ði = 1, 2, ..., nÞ, where dih is the time delay required by the agent place APi with respect to the hth color aih, denoted by dih = Dðaih Þ. 5. A  I ∪ O, which is the set of directed arc functions of agent place AP    and communication transition TC, where I APi , TCj aix , bjy :      CðLAPi Þ × C TCj ! N is the input function, and O APi , TCj aix , bjy :   CðLAPi Þ × C TIj ! N is the output function. 6. KB is the set of descriptive and control/decision-making knowledge with respect to agent place AP. KB :KBðAPi Þ − > SAPi is the mapping from agent place APi to the descriptive and control/decision-making knowledge set SAPi . 7. M0 = ½MðAPi Þn*1 , which is the initial mark of the AOKCTPN model, where MðAPi Þ :APS ! CðAPi Þ is the mark of APi , and i = 1, 2, ..., n. The interior behavior of the basic agent place is encapsulated in an AOKCTPN model. A basic agent place LAPj is defined by a 7-tuple:   LAPj = Pj , T j , Cj , Aj d , Aj inh , Kp , KT       Pj = LPj U MPj U SPj , which corresponds to the set of places in agent place LAPj, where LPj is a set of resource places to describe the interior state   of the agent; MPj = IMj , OMj is a set of message places to communicate between agent places, where IMj corresponds to an input message place from other agent to agent place LAPj and OMj corresponds to an output message place; SPj is a set of information place to describe the real-time information of an agent place.   2. T j = TI j , TSj , TEj , which is a set of delay time transitions, where TI j is the set of immediate transitions, TSj is a set of the stochastic transitions and TEj is a set of the deterministically timed transitions. In addition, ðLPj or     MPj or SPj Þ ∪ TI j or TSj or TEj ≠ Φ, ðLPj or MPj or SPj Þ ∪ TI j or TSj or TEj = Φ. 3. Cj corresponds to a color set of places and transitions in agent place       LAPj . Cj LPj , Cj MPj , Cj SPj , Cj ðTI j Þ and Cj ðTEj Þ corresponds to the respective color set of LPj , MPj , SPj , TI j , TSj andTEj . 4. Aj d is a set of connecting arc functions between places and transitions,     in which LPj or IMj or SPj × TI j or TSj or TEj ! f0, 1g is the input     function and TI j or TSj or TEj × LPj or OMj or SPj ! f0, 1g is the output function. 5. Aj inh is a set of inhibited arc functions, and Inhðpl , tm Þ : pl × tm ! f0, 1g is a inhibited arc function between place pl and transition tm. 6. Kp is the mapping from place set Pj to place knowledge set SPj . KT is a mapping from transition set Tj to transition knowledge set ST j . 1.

239

9.2 Agent-Oriented Knowledgeable Colored and Timed Petri Net

Table 9.1: The knowledge classification. Knowledge subset name

Detail description of the subset

Basic descriptive knowledge subset of places and transitions

1) State information of places and transitions. 2) The agent information that places and transitions belong to. 3) Token-type information corresponding to places and transitions. 4) Places’ capacity information. 5) Other information. 1) Token’s quantity statistical information in places. 2) Fired time statistical information of delay transitions and stochastic transitions. 3) Transitions’ fired frequency statistical information. 4) Other information. 1) The control and decision-making rules used to select tokens in places. 2) The control and decision-making rules used to solve conflicts among multi-transitions. 3) Other information.

Statistical descriptive knowledge subset of places and transitions

Control and decision-making knowledge subset of places and transitions

Communication transition

Message place

Immediate Stochastic transition transition

Information place

Directed arc Deterministic T01 transition

Resource place

Agent place

P11

Inhibited arc T02

Composite agent place

Figure 9.1: Places and transitions of an AOKCTPN.

The icons of places and transitions of the AOKCTON model are shown in Figure 9.1. The knowledge set of places and transitions can be classified into three subsets: basic descriptive knowledge subset, statistical descriptive knowledge subset, and control and decision-making knowledge subset. A detail description of the knowledge classification is shown in Table 9.1. 2.

Model operating rule The enabling and firing rules of the AOKCTPN model are presented as follows: Transition TCj is enabled with respect to the color bjs if MðAPi Þðair Þ ≥ IðAPi , TCj Þðair , bjs Þ, ∀APi 2 APS, air 2 CðAPi Þ, bjs 2 CðTCj Þ When TCj is fired, the new mark M′ becomes:

240

9 Scheduling Performance Evaluation

M′ðAPi Þðair Þ = MðAPi Þðair Þ + OðAPi , TCj Þðair , bjs Þ − IðAPi , TCj Þðair , bjs Þ

9.2.2 Modeling Process of an AOKCTPN Model 1.

Framework of an AOCKTPN-based AMHS model In order to reduce the complexity of the system modeling process and increase the reusability of the system modeling process, agents with the similar structure and function are aggregated to form the layered structure, which includes a basic agent layer, a subsystem layer and a system layer, as shown in Figure 9.2.

2.

Basic agent layer modeling As the basic elements of an AMHS model, the agents in the basic agent layer include the physical agent and the logical one. Corresponding to the manufacturing resources of AMHSs and SWFSs, the physical agents include a machine agent, a transportation rail agent, a shortcut agent, a turntable agent, and a stocker agent and so on. The logical agents with dispatching and control functions include a lot release agent, an AMHS scheduling agent, a routing agent (included in turntable agent), a zone control agent (included in transportation rail agent), etc. (1) Machine agent model The machine agent is the core in the basic agent layer, it is used to describe the characteristics of processing machines, wafer lot loading process, and wafer lot unloading process and so on. The AOKCTPN-based machine agent model is developed and shown in Figure 9.3. The detailed reasoning process of the machine agent model is described as follows: the token in the input message place IMi1 represents the lot-processing-request message sent by the

AMHS model

System layer

Interbay subsystem model

Subsystem layer

Material Basic handling agent scheduling layer agent model

Route control agent model

Stocker agent model

...

Transport Shortcut Turntable -ation rail agent agent agent model model model

Intrabay subsystem model

Material Machine handling Stocker agent scheduling agent model agent model model

Transport -ation rail agent model

Order release agent model

Figure 9.2: Architecture of an AOCKTPN-based scheduling performance evaluation model in AMHSs.

9.2 Agent-Oriented Knowledgeable Colored and Timed Petri Net

241

MReady_p Wait_te WaitS_c WaitS_ti

Enter_ti IMi1

Load_w

Processing_p

Load_te

Unload_w Unload_te

OK_te

OSi1

Outport_w

Exit_ti OMi1

TransportP_ti

OSi2

ISi1

OSi3

Figure 9.3: An AOKCTPN-based machine agent model.

transportation rail agent. The token in the place MReady p corresponds to machine idleness. When the machine is idle and receives the lot-processingrequest message, the immediate transition Enter ti is fired, the lot token enters the place Load w. When the lot is loaded by an idle machine, then the load delay transition Load te is fired, the lot enters the place Processing p. When the machine finished the operation, the delay transition OK te is fired and the lot enters the place Unload w. When the lot is unloaded by a machine, the unload delay transition Unload te is fired, the machine becomes idle and the machine token enters MReady p, and the lot token in place Unload w enters the place Output w. If the place Output w has token and information place ISi1 received information token, then the transition Exit ti is fired, and the lot token in place Output w enters the output message place OMi1 . Otherwise, the immediate transition WaitS ti is fired, the lot token enters the place WaitS c. After a set unit time tstand , the delay transition Wait te is fired, the lot token reenters the place Output w. The data types of different color are shown in Table 9.2. The color sets of the machine agent model are shown in Table 9.3. The knowledge set of the machine agent model is described in Table 9.4. In the machine agent model, the information place set fOSi1 , OSi2 , OSi3 g is used to output real-time state information of machines, and the information place ISi1 is used to input vehicles’ state information. (2)

Transportation rail agent model The AOKCTPN-based transportation rail agent model including a loadingsegment transportation rail agent model and an unloading-segment transportation rail agent model is developed and shown in Figure 9.4. The detailed reasoning process of the unloading-segment transportation rail agent model is described as follows: the token in the input message place IMi1 represents

242

9 Scheduling Performance Evaluation

Table 9.2: The types of color data. Data type name

Interpretation

Elements

LI P

Lot no. Product ID

S

Lot stage

PT

Product processing time

KM

Machine no.

E BAY

Machine state Material handling subsystem

PR

Transportation rail no.

PRT

Moving times of vehicles on transportation rail Transportation rail state Vehicle no.

Natural numbers   P1 , . . . , Pj , . . . , Pm   S1 , . . . , Sj , . . . , Sn , Send   PT1 , . . . , PTj , . . . , PTn   KM1 , . . . , KMj , . . . , KMn few , ee g fBay1 , . . . , Bayk , . . . , Bayn , Bayinter g   PR1 , . . . , PRj , . . . , PRn   PRT1 , . . . , PRTj , . . . , PRTn

PE VT VE ST STE TT TTE STIN/STOUT

fpw , pe g   VT1 , . . . , VTj , . . . , VTm Vehicle state fvw , ve g   Stocker no. ST1 , . . . , STj , . . . , STn Stocker state fsw , se g   Turntable no. TT1 , . . . , TTj , . . . , TTn Turntable state ftw , te g The time that lot enters/leaves system f0, tstand , . . . , k tstand , . . . , q tstand . . .g

Table 9.3: The color set of an AOKCTPN-based machine agent model. Name

Function interpretation

IMi1

MReady_p Enter_ti

Input message of the lot waiting for entering machine Output message of the lot leaving a machine Machine being idle Lot enters machine

Load_w

Prepare to load lots

Load_te

Lot be loaded by machines

Processing_p

Machine is processing

OK_te

Machine finishes process

OMi1

Unload_w

Prepare to unload lots

Unload_te

Lot be unloaded by machines

Outport_w

Lot being in machine’s outputport Lot is preparing to leave a machine Vehicle has been prepared to carry a lot

Exit_ti TransportP_ti

Color set   LI × fPi g × Sj × KM   LI × fPi g × Sj + 1 KM × fee g   LI × fPi g × Sj × KM × fee g   LI × fPi g × Sj × KM × fee g   LI × fPi g × Sj × KM × few g     LI × fPi g × Sj × KM × few g × PTj   LI × fPi g × Sj × KM × few g   LI × fPi g × Sj × KM × few g   LI × fPi g × Sj × KM × few g   LI × fPi g × Sj + 1 × KM × fee g   PR × fve g × LI × fPi g × Sj + 1 PR × fve g

9.2 Agent-Oriented Knowledgeable Colored and Timed Petri Net

243

Table 9.3: (continued ) Name

Function interpretation

Wait_te

Lot has been waiting for a unit time Lot is waiting vehicles

WaitS_c WaitS_ti

Lot starts to wait vehicles

OSi1

Output information “machine starts to process” Output information “machine finishes process” Output information “lot prepares to leave machine” Input information “vehicle is waiting to carry lot”

OSi2 OSi3 ISi1

Color set   LI × fPi g × Sj + 1   LI × fPi g × Sj + 1   LI × fPi g × Sj + 1

  KM × few g × LI × fPi g × Sj   KM × fee g × LI × fPi g × Sj   LI × fPi g × Sj + 1 PR × fve g

Table 9.4: The knowledge set of an AOKCTPN-based machine agent model. Name

Knowledge type

Knowledge description

Wait_te

Basic descriptive knowledge of transitions

Transition name: Wait te. Transition type: delay transition. Corresponding agent: Machine agent PM ap i. Delay time of transition: tstand (tstand is the set unit time of an AMHS system). The time of transition fired: Nim .

Outport_w

Statistics descriptive knowledge of transitions Basic descriptive knowledge of places

Control and decisionmaking knowledge of place

Place name: Outport w. Place type: resource place. Corresponding agent: machine agent PM ap i. Place capacity: 1. Rule 1: If the current place has token and information place ISi1 received information “a vehicle is waiting to carry a lot”, then transition Exit ti is fired. Rule 2: If the current place has token and information place ISi1 didn’t receive any information, then transition WaitS ti is fired.

the vehicle-entrance-request message sent by the upstream transportation rail agent. The token in the place Ready p corresponds to unloading-segment transportation rail idleness. When the unloading-segment transportation rail is idle and receives the vehicle-entrance-request message, the immediate

244

9 Scheduling Performance Evaluation

UWait_te UWaitS_c UWaitS_ti

Ready_p

NUnload_ti

Unload_p

KPoint1_p

ULoad_c OMi1

Unload_te IMi1

Run2_te

Run1_te

PEnter_ti

Kpoint2_p PLeave_ti

NRail1_ti

(a)

OMi2

OSi2

OSi1

Load_p

LPEnter_ti

Run4_te

NLoad_ti TStop_ti

IMi2

Load_te

Run3_te

OSi4

OSi5

OMi3 Load_c

Kpoint4_p LPLeave_ti

NRail2_ti

TStop_w

(b)

OSi3

LWait_te LWaitS_c LWaitS_ti

LReady_p

KPoint3_p

ISi1

IMi3

OSi6

ISi2

OSi7

Figure 9.4: An AOKCTPN-based transportation rail agent model: (a) a transportation rail agent model of vehicle unloading positions and (b) a transportation rail agent model of vehicle loading positions.

transition PEnter ti is fired, the vehicle token enters the place KPoint1 p, and then the delay transition Run1 te is fired. After a period of moving time PRTj , the vehicle token enters place Unload p. If the color of the vehicle token in place Unload p is fvw g × fKMik or STik g(KMik or STik is the machine or stocker No. of the current transportation rail connected), then the delay transition Unload te is fired. After a specific vehicle unloading time, the vehicle token enters place Uload c, and the lot token enters the output message place OMi2 . Otherwise, the immediate transition NUnload ti is fired, and the vehicle token enters place Uload c directly. When Uload c received the vehicle token, the delay transition Run2 te is fired. After a specific vehicle moving time, the vehicle arrives the terminal position of the current transportation rail and the vehicle token enters the place Kpoint2 p. If the place Kpoint2 p has vehicle token and information place ISi1 received information token, then the immediate transition PLeave ti is fired, and the vehicle token in place Kpoint2 p enters the output message place OMi1 . Otherwise, the immediate

9.2 Agent-Oriented Knowledgeable Colored and Timed Petri Net

245

transition UWaitS ti is fired, the vehicle token enters place UWaitS c. After a set unit time tstand , the delay transition UWait te is fired, the vehicle token re-enters the place Kpoint2 p. The reasoning process of the loading-segment transportation rail agent model is similar to that of the unloading-segment transportation rail agent model, due to space limitations, the reasoning process of the loading-segment is not presented in this section. (3)

Shortcut agent model The AOKCTPN-based shortcut agent model is developed and shown in Figure 9.5. The detailed reasoning process of the shortcut agent model is described as follows: the token in the input message place IMi1 represents the vehicle-entrance-request message sent by the upstream turntable agent. The token in the place SReady p corresponds to shortcut idleness. When the shortcut is idle and receives the vehicle-entrance-request message, the immediate transition SEnter ti is fired, the vehicle token enters the place SPoint1 p, and then the delay transition SRun1 te is fired. After a specific moving time, the vehicle token enters place Spoint2 p, If the place Spoint2 p has vehicle token and information place ISi1 received information token, then the immediate transition SLeave ti is fired, and the vehicle token in place Spoint2 p enters the output message place OMi1 . Otherwise, the immediate transition SWaitS ti is fired, the vehicle token enters place SWaitS c. After a set unit time tstand , the delay transition SWait te is fired, the vehicle token re-enters the place Spoint2 p.

(4)

Stocker agent model The AOKCTPN-based stocker agent model is built up and shown in Figure 9.6. The detailed reasoning process of the stocker agent model is described

SReady_p SWaitS_c

SWait_te IMi1

SEnter_ti

SWaitS_ti

SPoint1_p

Spoint2_p

SLeave_ti

SRun1_te

OMi1

NTtable_ti

OSi1

OSi2

ISi1

Figure 9.5: An AOKCTPN-based shortcut agent model.

OSi3

246

9 Scheduling Performance Evaluation

BWait_te

BWaitS_ti

BWaitS_c BEnter1_ti Bport1_w

Stocker_in

BExit1_ti BPort2_w

BLoad_te

n BUnload_te

IMi1

BExit2_ti

OMi1

NRail1_ti IMi2

OMi2

BEnter2_ti

OSi1

OSi2

ISi1

OSi3 OSi4

Figure 9.6: An AOKCTPN-based stocker agent model.

as follows: the token in the input message place IMi1 and IMi2 represents the vehicle-entrance-request message sent by the transportation rail agent of an Intrabay system and an Interbay system, respectively. In the process of restoring wafer lots, when the input message place IMi1 or IMi2 in the stocker agent receives the wafer lot entrance-request message, the immediate transition BEnter1 ti or BEnter2 ti is fired, the wafer lot token enters the place Bport1 w. After a stochastic storing time (i.e., stochastic transition BUnload te is fired), the wafer lot token enters the place Stocker in. When a quantity of wafer lot token in place Stocker in equals to the capacity of the stocker, the immediate transitions BEnter1 ti and BEnter2 ti are prohibited to be fired. In the process of pulling out wafer lots, when the place Stocker in has wafer lot tokens, which one wafer lot be taken out firstly is decided based on the knowledge of the place Stocker in (e.g., highest priority rule, longest wait time rule and so on). After the wafer lot token is selected, then stochastic transition BLoad te is fired and the wafer lot token enters place BPort2 w. If the place BPort2 w has wafer lot token and information place ISi1 received information “vehicle in the Intrabay system prepares to carry wafer lot”, then the immediate transition BExit1 ti is fired, and the wafer lot token in place BPort2 w enters the output message place OMi1 . Else if the place BPort2 w has wafer lot token and information place ISi1 received information “vehicle in the Interbay system prepares to carry wafer lot”, then the immediate transition BExit2 ti is fired, and the wafer lot token enters the output message place OMi2 . Otherwise, the immediate transition BWaitS ti is fired, the vehicle token enters the place BWaitS c. After a set unit time tstand , the delay transition BWait te is fired, the vehicle token re-enters the place BPort2 w.

247

9.2 Agent-Oriented Knowledgeable Colored and Timed Petri Net

OSi5

OSi3

TReset1_te

TBack1_p

IMi2

OSi6 Tpoint5_p

TEnter2_ti

TReady1_p KPoint1_p

IMi1

TEnter1_ti

OMi3 TLeave1_ti TLeave1_ti

OSi1

ISi1

Kpoint2_p

Tpoint3_p

NRail1_ti Turn1_te

(a)

OMi1

OMi2 Tleave2_ti

OSi2

IMi3

(b)

TPoint6_p Turn2_te

TBack_te

OSi4

Figure 9.7: An AOKCTPN-based turntable agent model: (a) model with single-input and multi-output interface and (b) model with multi-input and single-output interface.

(5)

Turntable agent model The AOKCTPN-based turntable agent model including a turntable agent model with single-input and a multi-output (SIMO) interface and turntable agent model with multi-input and single-output (MISO) interface is built up and shown in Figure 9.7. The detailed reasoning process of SIMO turntable agent model is described as follows: the token in the input message place IMi1 represents the vehicle-entrance-request message sent by the upstream transportation rail agent. The token in the place TReady1 p corresponds to turntable idleness. When the turntable is idle and receives the vehicleentrance-request message, the immediate transition TEnter1 ti is fired, and vehicle token enters the place KPoint1 p. If the place KPoint1 p has vehicle token and information place ISi1 received information token, then the delay transition Turn1 te and immediate transition Tleave2 ti are fired in sequence. The vehicle token in place KPoint1 p enters the place OMi2 . Otherwise, the immediate transition TLeave1 ti is fired and the vehicle token enters the place OMi1 , that is, the vehicle enters the downstream transportation rail. Since the reasoning processes of the MISO turntable agent model is similar to that of the SIMO turntable, the detailed reasoning process of the MISO turntable is not introduced in this section.

(6)

Order release agent model Figure 9.8 shows the order release agent model based on an AOKCTPN. The order release process is as follows: the token in Release_s indicates that the system is ready to start release orders. Initially, the token in the place Release_s triggers the instantaneous transition place Release_p and then triggers the instantaneous transition place Exit_ti. The wafer leaves the order release agent model and enters the feeding output processing place

248

9 Scheduling Performance Evaluation

Release_s

Release_ti

Release_p

Exit_ti

OMi1

Interval3_te Interval2_ts

Interval_t



Interval1_ti ISi1 Figure 9.8: An AOKCTPN-based order release agent model.

IMi1

OMi1 Enterkin_ti Send_ti

ISkin

Receive_ti OSkout

Send_p

Exit_ti

ISi

OSj Enteri_ti

IS2

Dispatch_ti

Receive_p

OS2 Enter2_ti

IS1 Enter1_ti

OS1

Figure 9.9: An AOKCTPN-based material transportation scheduling agent model.

OMi1. At the same time, the status bit Interval_t is used to trigger the transition time bit of the response according to the material setting information. The random delay transition time Interval2_ts corresponds to the stochastic distribution order release strategy, which determines the delay time interval Interval3_te corresponding to the fixed time interval of the order release strategy. (7)

Material handling scheduling agent model The material handling scheduling agent model based on AOKCTPN method is shown in Figure 9.9. The material handling scheduling process is as follows: when the transport cart finishes unloading the wafer card or the wafer card arrives at the processing equipment output port for transportation, and the corresponding information place OSi receives the material

9.2 Agent-Oriented Knowledgeable Colored and Timed Petri Net

249

handling scheduling request, triggers the instantaneous transition position Enteri_ti and enters the material handling scheduling optimization request status place Send_p. The status bit Send_p triggers the transient transition place Sent_ti to send a material handling scheduling request to the realtime optimization dispatch module. When the real-time optimization scheduling module forms the optimal scheduling scheme, the material handling scheduling agent model receives the optimal scheduling scheme through the transient transition Receive_ti. After the transient transition Exit_ti, the optimal material handling scheduling information is sent out. 3.

Subsystem layer modeling Based on several basic agent layer models and communication transitions, the subsystem layer model can be constructed. Since AMHSs contain Interbay subsystems and Intrabay subsystems, the subsystem layer model can be subclassified as an Interbay subsystem model and an Intrabay subsystem model. (1) Intrabay subsystem model The AOKCTPN-based Intrabay subsystem model is developed and shown in Figure 9.10. The detailed reasoning process of the Intrabay subsystem model is described as follows: the token in the input message place IMi1 corresponds to the wafer lot requesting for entering the Intrabay subsystem. By firing the communication transition SP ti, the wafer lot token enters the stocker agent place Stocker ap. When the wafer lot in the stocker is preparing to move to the machine corresponding to the current processing step (suppose that the machine agent is PM ap i), the optimal empty vehicle is designated to carry it based on the Intrabay’s intelligent scheduling approach. When the designated empty vehicle token enters the transportation rail agent VPA ap n + 1, the communication transitions RK ti n + 1 and TK ti n + 1 are fired, and the empty vehicle token in VPA ap n + 1 and the wafer lot token in Stocker ap enter the transportation rail VPB ap n + 1. The wafer in the stocker is loaded into the empty vehicle. After firing the communication transition RR ti n + 1 and CR ti, the occupied vehicle token enters the route agent place Route ap. Guided by the running path table of the Intrabay system, occupied vehicle token chooses to enter the downstream transportation rail in sequence. When the occupied vehicle token enters the transportation rail agent VPA ap i, the communication transitions TP ti i and RK ti i are fired concurrently. The wafer lot token in occupied vehicle is unloaded to machine agent PM ap i and the emptied vehicle token enters the downstream transportation rail agent VPB ap i. When the wafer lot in machine agent PM ap i is processed, then the optimal empty vehicle is designated to carry the wafer lot based on the Intrabay’s intelligent scheduling approach again. After loading and transporting by designated vehicles, the wafer lot token is moved to the

250

9 Scheduling Performance Evaluation

CR_ti

RB_ti_1

RR_ti_1

RK_ti_1 VPA_ap_1

TP_ti_1

TK_ti_1

VPB_ap_1

PM_ap_1 ...... RB_ti_n

RR_ti_n RK_ti_n

Route_ap VPA_ap_n

TP_ti_n

TK_ti_n

VPB_ap_n Recyle_ap

PM_ap_n RB_ti_n+1

Stocker_ap

VPA_ap_n+1

TP_ti_n+1 IMi1

RR_ti_n+1

RK_ti_n+1

SP_ti

VPB_ap_n+1

TK_ti_n+1 SK_ti

OMi1

Figure 9.10: The AOKCTPN-based Intrabay subsystem model.

stocker agent Stocker ap. After the wafer lot finished all process in the Intrabay subsystem, the communication transition SK ti is fired and the wafer lot token enters the output message place OMi1 , the wafer lot leaves the current Intrabay subsystem. (2)

Interbay subsystem model The AOKCTPN-based Interbay subsystem model is developed and shown in Figure 9.11. The detailed reasoning process of the Interbay subsystem model is described as follows: the token in the input message place IMi corresponds to the wafer lot requesting for entering the Interbay subsystem. By firing the communication transition RP ti i, the wafer lot token enters the stocker agent place Stocker ap i. When the wafer lot in a stocker requests to move to the destination stocker (suppose that the destination stocker is Stocker ap j), the optimal empty vehicle is designated to carry it based on the Interbay’s intelligent scheduling approach. When the designated empty

9.2 Agent-Oriented Knowledgeable Colored and Timed Petri Net

251

CR_ti

RR_ti_1

RB_ti_1 RK_ti_1 VPA_ap_1

SP_ti_1 Stocker_ap_1 SK_ti_1

RP_ti_1

IM1

VPB_ap_1

EK_ti_1

OM1

RR_ti_n

RB_ti_n RK_ti_n VPA_ap_n

SP_ti_n Stocker_ap_n SK_ti_n RP_ti_n

IMn RATB_ap_1

RBTB_ap_1

RATB_ap_m

EK_ti_n

ATTable_ap_1

OMn

RATR_ap_1

RTP_ap_1

Route _ap

VPB_ap_n

RTK_ap_1 Recyle_ap

BTTable_ap_1 RBTR_ap_1

Shortcut_ap_1

RATR_ap_m

ATTable_ap_m RTP_ap_m

RTK_ap_m BTTable_ap_m RBTR_ap_m

RBTB_ap_m Shortcut_ap_m

Figure 9.11: The AOKCTPN-based Interbay subsystem model.

vehicle token enters the transportation rail agent VPA ap i, the communication transitions RK ti i and SK ti i are fired, the empty vehicle token in VPA ap i and the wafer lot token in Stocker ap j enter the transportation rail place VPB ap i and the wafer in the stocker is loaded into an empty vehicle. After firing the communication transition RR ti i and CR ti, the occupied vehicle token enters the route agent place Route ap. Guided by running path table of the Interbay system, the occupied vehicle token chooses to enter the downstream transportation rail or the shortcut until arrives at the transportation rail VPA ap j. After the occupied vehicle token enters the transportation rail agent VPA ap j, the communication transitions SP ti j and RK ti j are fired. On the one hand, when the communication transition SP ti j is fired, the wafer lot token is unloaded to the stocker agent Stocker ap j. After firing the communication transition EK ti j, the wafer lot enters the output message place OMj . On the other hand, the

252

9 Scheduling Performance Evaluation

RBI_ti_1

Intrabay_ap_1

RBO_ti_1

RBI_ti_i

Intrabay_ap_i

RBO_ti_i

RBI_ti_n

Intrabay_ap_n

RBO_ti_n

Interbay_ap

Release_ap

RRI_ti

RRO_ti

End_ap

Figure 9.12: The AOKCTPN-based system layer model.

communication transition RK ti j is fired and the empty vehicle token enters the downstream transportation rail agent VPB ap j. The empty vehicle runs on the transportation rail continuously. 4.

System layer modeling The AOKCTPN-based system layer model is developed and shown in Figure 9.12. The detailed reasoning process of the system layer model is described as follows: according to controlling and decision-making knowledge, the release agent place Release ap generates a wafer lot token. After firing the communication transition RRI ti, the wafer lot token enters the composite agent place Interbay ap. Based on the wafer lot’s manufacturing process, the corresponding communication transition RBI ti i is fired and the wafer lot token enters the composite agent place Intrabay ap i. After finishing the current processing step in the composite agent place Intrabay ap i, the wafer lot token returns the composite agent place Interbay ap. If the wafer lot has finished all steps, the wafer lot token enters exit agent place End ap. Otherwise, the wafer lot token re-enters the Intrabay subsystem agent for the next processing step. The color sets of the system layer model are shown in Table 9.5. The knowledge sets of the system layer model are described in Table 9.6.

9.2.3 Feasibility Analysis of an AOKCTPN The feasibility analysis of an AOKCTPN model is used to qualitatively analyze the AOKCTPN model under the guidance of the high-level PN theory to test whether the AOKCTPN model’s structure and operation process are correct and feasible. The feasibility of the AOKCTPN model is the prerequisite and basis for the

9.2 Agent-Oriented Knowledgeable Colored and Timed Petri Net

253

Table 9.5: The color sets of an AOKCTPN-based system layer model. Name

Function interpretation

Color set

Release_ap RRI_ti Interbay_ap RBI_ti_i Intrabay_ap_i RBO_ti_i RRO_ti End_ap

Wafer lot release Wafer lot requests to enter an AMHS Wafer lot is running in an Interbay subsystem Wafer lot requests to enter an Intrabay subsystem Wafer lot is running in an Intrabay subsystem Wafer lot requests to enter an Interbay subsystem Wafer lot requests to leave an AMHS Wafer lot exits an AMHS

LI×{Pi}×{S1} LI×{Pi}×{S1} LI×{Pi}×{Sj+1}×{Bayinter} LI×{Pi}×{Sj}×{Bayk} LI×{Pi}×{Sj}×{Bayk} LI×{Pi}×{Sj+1}×{Bayinter} LI×{Pi}×{Send} LI×{Pi}×{Send}

Table 9.6: The knowledge sets of an AOKCTPN-based system layer model. Name

Knowledge type

Knowledge description

Release_ap

Basic descriptive knowledge of places Statistics descriptive knowledge of places Control and decisionmaking knowledge of places

Place name: Release ap. Place type: Agent place. The number of wafer lot that leaved place: r1i . Rule1: The token is generated based on the constant interval time. Rule 2: The token is generated based on the stochastic interval time. Rule 3: The token is generated based on the CONWIP (constant work-in-process) rule.

performance evaluation of the scheduling method for AMHSs. The main properties of the feasibility analysis of the model include boundedness and activity, that is, the model is structurally live and bounded. The concrete steps of the feasibility analysis of the AOKCTPN model are presented as follows: Step 1: Establish the AOKCTPN model of each Intrabay and Interbay subsystem based on the AOKCTPN modeling method. Step 2: Based on the regularization and simplification of the scheduling/control strategy and knowledge of each basic agent module, the agent bits and the compound proxy bits are expanded to the CTPN model to describe the internal behavior of the agent bit. Step 3: Considering the complexity and large-scale characteristics of a CTPN model, the CTPN model is simplified by using modularized extended CTPN model under the premise of keeping the model live and bounded in order to reduce the complexity of the model. Step 4: Establish the reachability map of a simplified CTPN model in order to check the activity and boundedness of the AOKCTPN model.

254

9 Scheduling Performance Evaluation

On the basis of the simplification theory of PN and the simplification method of the timed PN, the simplification rules of the modularized extended CTPN model are presented as follows: Definition 1: E = fP, T, C, I, O, M0 g is defined as the CTPN-based system model, n o T=fTI, TDg. EA = PA , T A , CA , I A , OA , M0A is defined as an agent module based on   the CTPN of system model E, T A = TI A , TDA , EA 2 E. Definition 2: S = ðM, FÞ is defined as a state of CTPN model E, where M is the state of model E, F is the set of firing time intervals of the model and LðEÞ is the transition triggering sequence set. Set t0 is trigger at the model initial time θ0 , the system state changes from S0 = ðM0 , F0 Þ to S1 = ðM1 , F1 Þ; at time θi the transition ti triggers the system state to Si + 1 = ðMi + 1 , Fi + 1 Þ. The transition triggering sequence is $ = ðt0 , θ0 Þðt1 , θ1 Þ...ðti , θi Þ. Rule 1: For the agent module EA shown in Figure 9.13, EA .Pin =f IP1 g, EA .Pout =f OP1 g.,∀M0 , MðPiA Þðaik Þ < IðPiA , TjA Þðaik , bjs Þ. If: (1) when IP1 receives a token, OP1 receives a token after T(0 < T < ∞); (2) when OP1 receives a token, IP1 can receive another token. Then the agent module EA can be simplified to EA ′ (Figure 9.13), where (1) IP1 ′  = OP1 ′ = f tA′ g, tA′  = f OP1 ′ g,  tA′ = f IP1 ′ g; (2) CðOP1 ′ Þ = CðOP1 Þ, CðIP1 ′ Þ = CðIP1 Þ, ts P CðtA′ Þ = Cð  OP1 Þ+f tdAi g, ts is the quantity of delay transitions in EA ; (3)





i=1

P

IðIP1 , t Þ = A′

P 2PA , T 2T A i

IðPi , Tj Þ, OðOP1 ′ , tA′ Þ =

i

P

OðPi , Tj Þ.

P 2PA , T 2T A i

i

Rule 2: For the agent module EA shown in Figure 9.14, EA .Pin =f IP1 g, EA .Pout =f OP1 , OP2 g. ∀M0 , MðPiA Þðaik Þ < IðPiA , TjA Þðaik , bjs Þ. If : (1) when IP1 receives a token, OP1 or OP2 receives a token after T (0 < T < ∞); (2) when OP1 or OP2 receives a token, IP1 can receive another token. Then the agent module EA can be simplified to EA ′ (Figure 9.14), where (1) IP1 ′  =   f t1A′ , t2A′ g,  OP1 ′ = f t1A′ g,  OP2 ′ = f t2A′ g, tA′ = fOP1 ′ g, tA′ = fOP2 ′ g,  t1A′ =  t2A′ = 1 2

P3 t A′ IP1

T1 P1

T2 P2

T3

OP1

IP1′

OP1′

Figure 9.13: Model simplification rule 1.

255

9.2 Agent-Oriented Knowledgeable Colored and Timed Petri Net

IP1

T1 P1

T4

P3

T5

T2

P2

T3

OP1

IP1′

t A2′

OP2

IP1

P1

T1

T4 P3

T5

IP1′ OP1

IP2 T2

P2

t A1 ′

OP1′

OP2′ Figure 9.14: Model simplification rule 2.

t A1 ′

IP2′

OP1′ t A2′

T3

Figure 9.15: Model simplification rule 3.

f IP1 ′ g; (2) Cð OP1 ′ Þ = CðOP1 Þ, CðOP2 ′ Þ = Cð OP2 Þ, CðIP1 ′ Þ = CðIP1 Þ, Cðt1A′ Þ = P P tdAi g, Cðt2A′ Þ = Cð  OP2 Þ+f tdAj g, ;1 and ;2 are the sets of Cð  OP1 Þ+f tdA 2;1

tdA 2;2

i

j

locations and transitions sequences in EA ; (3) IðIP1 ′ , tA′ Þ = OðOP1 ′ , tA′ Þ = P Pi , Tj 2;2

P Pi , Tj 2S1

OðPi , Tj Þ, IðIP2 ′ , tA′ Þ =

P Pi , Tj 2;2

P Pi , Tj 2;1

IðPi , Tj Þ,

IðPi , Tj Þ, OðOP2 ′ , tA′ Þ =

OðPi , Tj Þ.

Rule 3: For the agent module EA shown in Figure 9.15, EA .Pin = f IP1 , IP2 g,EA .Pout = f OP1 g. ∀M0 , MðPiA Þðaik Þ < IðPiA , TjA Þðaik , bjs Þ. If: (1) when IP1 or IP2 receives a token, OP1 receives a token after T(0 < T < ∞); (2) when OP1 receives a token, IP1 or IP2 can receive another token. Then the agent module EA can be simplified to EA ′ (Figure 9.15), where (1)  OP1 ′ =   f t1A′ , t2A′ g, IP1 ′ = f t1A′ g, IP2 ′ = f t2A′ g,tA′ = f OP1 ′ g,tA′ = f OP1 ′ g,  t1A′ = f IP1 ′ g,  t2A′ 1 2

= f IP2 ′ g; (2) Cð OP1 ′ Þ = Cð OP1 Þ,CðIP1 ′ Þ = Cð IP1 Þ, CðIP2 ′ Þ = CðIP2 Þ,Cðt1A′ Þ = Cð  OP1 Þ P P tdAi g,Cðt2A′ Þ = Cð  OP1 Þ+f tdAj g, S1 and S2 are the sets of locations +f tdA 2;1

tdA 2;2

i

j

and transitions sequences in EA when IP1 ! OP1 and IP2 ! OP1 ; (3) IðIP1 ′ , t1A′ Þ P P P IðPi , Tj Þ,IðIP2 ′ , t2A′ Þ = IðPi , Tj Þ,OðOP1 ′ , t1A′ Þ = OðPi , Tj Þ, = Pi , Tj 2;1 ′

OðOP1 ,

t2A′ Þ

=

P Pi , Tj 2;2

Pi , Tj 2;2

OðPi , Tj Þ.

Pi , Tj 2;1

256

9 Scheduling Performance Evaluation

Rule 4: For the agent module EA shown in Figure 9.16, EA .Pin = f IP1 g, EA .Pout = f OP1 g.∀M0 , MðPiA Þðaik Þ < IðPiA , TjA Þðaik , bjs Þ. If: (1) when IP1 receives a token, OP1 receives a token after T(0 < T < ∞); (2) when OP1 receives a token, IP1 can receive another token. Then the agent module EA can be simplified to EA′ (Figure 9.16), where: (1) IP1 ′  =  OP1 ′ = f tA′ g, tA′  = fOP1 ′ g,  tA′ = f IP1 ′ g; (2) CðOP1 ′ Þ = Cð OP1 Þ,CðIP1 ′ Þ  ts  P = Cð IP1 Þ,CðtA′ Þ = Cð  OP1 Þ+ wi  tdAi , ts is the quantity of delay transitions i=1

in EA ,wi is the number of triggers of delay transitions tdAi in EA ; (3) IðIP1 ′ , tA′ Þ P P IðPi , Tj Þ,OðOP1 ′ , tA′ Þ = OðPi , Tj Þ. = P 2PA , T 2T A i

P 2PA , T 2T A

i

i

i

Rule 5: For the agent module EA shown in Figure 9.17, EA .Pin = f IP1 g,EA .Pout = f OP1 g。 ∀M0 , MðPiA Þðaik Þ < IðPiA , TjA Þðaik , bjs Þ. If: (1) when IP1 receives a token, OP1 receives a token after T(0 < T < ∞); (2) when OP1 receives a token, IP1 can receive another token. Then the agent module EA can be simplified to EA ′ (Figure 9.17), where (1) IP1 ′  =  OP1 ′ = f tA′ g,tA′  = fOP1 ′ g,  tA′ = f IP1 ′ g;(2)Cð OP1 ′ Þ = CðOP1 Þ,Cð IP1 ′ Þ ts ts A  A P P = Cð IP1 Þ,CðtA′ Þ = Cð  OP1 Þ+ tsum , tsum 2 ½min tdAi , max tdAi , ts is the numi=1

P

ber of delay transitions in EA ; (3) IðIP1 ′ , tA′ Þ =

P 2PA , T 2T A i i

P

=

P 2PA , T 2T A i

OðPi , Tj Þ。

T3 P2

IP1

IðPi , Tj Þ,OðOP1 ′ , tA′ Þ

i

T4

IP1

i=1

T1

T1 T2 T3

t A′

P1

T2

P1

T4

OP1

IP1′

OP1′

Figure 9.16: Model simplification rule 4.

t A′ OP1 IP1′

OP1′ Figure 9.17: Model simplification rule 5.

9.3 Scheduling Performance Evaluation of an AMHS Based on the AOKCTPN

257

9.3 Scheduling Performance Evaluation of an AMHS Based on the AOKCTPN The evaluation indicators of the scheduling performance of an AMHS consist of two aspects: performance indicators of the material handling system and performance indicators of the processing system. The former one includes vehicle’s transportation quantity, wafer lot’s waiting time, vehicle’s transport time, wafer lot’s delivery time and vehicle utilization. The latter one includes wafer lot’s cycle time, throughput, due date satisfactory rate and WIP quantity. Among them, wafer lot’s waiting time, transport time, delivery time, cycle time and due date satisfactory rate can be determined by the statistical analysis of the transition time of an AOKCTPN model, and vehicle’s transportation amount, utilization, throughput and WIP quantity can be determined by the trigger times of transient transitions and delayed transitions.

9.3.1 Transition Time There are two kinds of transitions in an AOKCTPN model: transient transition and delayed transition. The transient transition is triggered immediately when the trigger conditions are satisfied, no time-consuming, such as the status of the wafer lot changes after all processing steps are completed, transitions of communication. The delayed transition takes some time to be triggered after the trigger conditions are satisfied, such as a vehicle loads/unloads a wafer lot. Therefore, an AOKCTPN should confirm all the time parameters of the delayed transition in order to obtain the statistics of the material handling system. For the normal timed PN, the delayed transition can be described by a definite time parameter. For example, in Figure 9.18 (a), transition Tj1 and transition Tj2 denote that wafer lots 1 and 2 are processed on

t=3

t=7 Tj1

t = [3(Black), 7(Gray)]

Tj2

Tk1

Black tokens Ti1 Equipment processing wafer card no.n1 Tj2 Equipment processing wafer card no.n2

(a)

Ti1

t = Pλ ∈[tinf , tsup]

Gray tokens

Tk 1 Equipment processing wafer card The black token is wafer card no.n1, the gray tokens is wafer card no.n2

(b)



Ti1 Wafer card loading/unloading time in a storage warehouse The black token is the wafer card no.n1

(c)

Figure 9.18: Determination of the delayed transition time parameters in an AOCKTPN: (a) definition of the transition time in a classical PN; (b) definition of the delayed transition time in an AOKCTPN; and (c) definition of the random transition time in an AOKCTPN.

258

9 Scheduling Performance Evaluation

machine 1, and the time parameters of Tj1 and Tj2 are the processing times of wafer lots 1 and 2, which are 3 and 7, respectively. For an AOKCTPN model, since it describes the processing procedure by using only a transition, the delay time is different when tokens representing different wafer lots trigger transitions, that is to say, the parameter of delayed transition is not unique. Thus, it is required to determine the appropriate time parameters according to taken colors of each resource in the model. As shown in Figure 9.18(b), transition Tk1 denotes that machine 1, and wafer lots 1 and 2 are, respectively, represented by black and gray tokens, whose time parameters are defined as [3(black), 7(gray)]. For the existence and influence of various uncertainties in the AMHS, they are described by the random delayed transitions and the sending times of all the delayed transitions are random variables Pλ subject to a certain possibility distribution Pλ 2 ½tinf , tsup , as shown in Figure 9.18 (c). For example, the random delayed transition is that wafer lot is loaded and unloaded in the stocker.

9.3.2 Scheduling Performance Evaluation Indicators The performance indicators of an AMHS and a wafer processing system include vehicle movement M, wafer lot’s waiting time WT, WIP (work in process), wafer lot’s transport time TT, wafer lot’s delivery time DT, vehicle utilization ηvehicle , throughput T, wafer lot’s cycle time CT and wafer lot’s due date satisfactory rate DS. And the AOKCTPN-based performance indicators are determined by using eqs (9.1–9.9). M=

Rt X

i Nu3

(9:1)

i=1

WT =

Mt X

i Nm tstand

+

i=1

 Rt P i=1

+

=M

(9:2)

St X ðNsiin − Nsiout Þ

(9:3)

i=1

½Nl1i tRun3 Rt P i=1

+

! j Nbw tstand

j=1

WIP =

TT =

Bt X

St P j=1

te

+ Nl3i tLoad

i ½Nu1 tRun1

j ðNs1 ðtSRun1

te

te

te

+ ðNl1i + Nl3i ÞtRun4

i + Nu3 tLoad

te

te

i + Nlw1 tLWait te 

i i + ðNu1 + Nu3 ÞtRun2

+ tSWait te ÞÞ +

St P j=1

te

i + Nuw1 tUWait te 

(9:4)

 j ðNAt1 tTurn1

DT = TT + WT

te

j + NBt1 tTBack te Þ =M

(9:5)

9.3 Scheduling Performance Evaluation of an AMHS Based on the AOKCTPN

 ηvehicle = ðTT*MÞ= +

Rt P i=1

Rt P i=1

½Nli ðtRun3

i ½Nu1 ðtRun1

ðtSRun1

te

te

te

259

i + tRun4 te Þ + Nl3i tLoad te Nlw tLWait te  +

i + tRun2 te Þ + Nu3 tLoad

+ tSWait te ÞÞ +

St P j=1

T=

j ðNAt tTurn1

St X

St P i + Nuw tUWait te  + ðNsj j=1  j te + NBt tTBack te Þ

te

Nsiout

(9:6)

(9:7)

i=1

CT =

T X i=1

DS =

T X i=1

i ðCTlot Þ=T =

T X i i ðSTOUT − STIN Þ=T

(9:8)

i=1

8